JPA Automation
Five years of MEP delivery, with an automation programme running alongside.
How the practice was run
Bidding & wins
Pursued and won framework places and individual commissions. Tender response, fee negotiation, scope definition, programme commitment.
Contract negotiation
Appointment terms, professional indemnity exposure, scope of services, fee structure, and change-management mechanisms across consultant and sub-consultant agreements.
Design assurance
Owned the technical-review chain. Drawings, calculations, schedules, and specifications signed off against BS EN, CIBSE, and UK Building Regulations before issue.
Compliance & disputes
Statutory compliance correspondence, contractor RFIs, design queries, and dispute resolution. Involved at every step where the practice carried regulatory or contractual exposure.
P&L ownership
Forecasting, invoicing, cash flow, resource planning, and margin discipline across the project portfolio. Direct line into commercial decisions.
Client relationships
Direct point of contact for principal contractors, developers, and main contractor design teams. Long-term framework relationships kept warm through delivery quality.
Automation programme
01
Diagnosis
Practice was profitable but unscalable. Engineering bottlenecks at calculation, specification, and reporting absorbed 60%+ of capacity. Margin trapped at 30% despite a strong client base.
02
Architecture
Built deterministic tooling first (calculation engines, specification generators, scheduling automations) before introducing generative components. Auditability over flexibility for regulated work.
03
Operationalisation
Repetitive engineering work was replaced with reviewable system outputs. Engineers shifted from production to design assurance and client engagement. The team settled at four.
04
Outcome
~£1m annual turnover sustained. Gross margin moved to 70%+. Frameworks delivered for a Tier 1 tech sector client (NDA), Ministry of Justice, Ministry of Defence, Willmott Dixon, LJJ, Fit Out UK.
Commercial-side automation · the other half of the programme
Automated SEO
Improved local-search coverage for target MEP-consultant queries across selected geographies. The content programme was tied to project completions — every closed job seeded a new case study with schema markup, OG metadata, and internal linking generated from the project record.
- First-page rankings for the target keyword set
- Schema + sitemap regenerated on every release
- Local listings kept in sync across Google, Bing, and Apple
- Backlink + Core Web Vitals monitoring in CI
Automated inbound response
SEO did the work of getting the right people onto the site. From there, an Outlook MCP server read each incoming enquiry, parsed it against the project-scoping outline for that service line, and drafted a reply that addressed only the information that was missing — not a templated brochure dump. Quote-ready conversations reached an engineer warm; vague ones came back with one targeted ask, not a back-and-forth.
- Outlook MCP server reading inbound threads in real time
- Parsed against the per-service scoping outline (scope, site, programme, budget, deliverables)
- Reply drafted to address only the gaps — sent on engineer review, never auto-send
- First-response latency dropped from days to minutes without losing the human signature
544
Projects directly involved with
30 → 70%
Gross margin
17 → 4
Engineers, ~£1m turnover
424
C# files, AutoCAD–Revit bridge
Before / after
2018 · pre-automation
Engineers
17
Gross margin
30%
Cycle
calc → redraw → respec → reschedule → re-report
Engineering hours
60%+ absorbed by production
2025 · post-automation
Engineers
4
Gross margin
70%+
Cycle
model → review → issue
Engineering hours
design assurance + client
Hours per deliverable · before / after
Cable schedule
MEP spec section
Condition report
Calc set (sample)
Indicative timings drawn from the four highest-frequency deliverables. Engineer time saved was redirected to design assurance and client engagement — the practice did not shrink, it shifted weight.
Frameworks delivered
Automated MEP Workflows and Calculations
424 production C# files. The technical spine of the practice.
The technical spine of the practice. A unified .NET codebase that replaced the calculate-redraw-respecify-rereport cycle every MEP consultancy carries. Calculations auto-populate from the live Revit and AutoCAD models and write their results back to the same families and schedules — a closed round-trip that reduced manual rekeying between calculations, drawings, schedules, and reports.
Live sample calculations
drawing ⇄ calc ⇄ drawing · no rekeying, no drift
Sample calculation
LPHW flow — diameter from heat load
Source
Revit · Space + Mechanical Equipment
Calc engine
deterministic · cited · versioned
Target
Revit · Pipe Type + Pipe Segment
Inputs auto-populate from the live drawing · results write back to the same families and schedules · single source of truth with zero information loss.
Heat load
← Revit · Space.HeatingDesignLoad
Loop ΔT
← Revit · MechSystem.DeltaT_K
Water density
← Project · Fluid Library @ 70 °C
Specific heat
← Project · Fluid Library @ 70 °C
Max velocity
← Standards · CIBSE Guide C §4.1
Mass flow ṁ
Q·1000 / (cp·ΔT)
Volume flow V̇
ṁ / ρ × 1000
Required area
V̇ / vmax
Required bore
√(4A/π)
Pipe size
round-up next listed DN
→ Revit · PipeType.NominalDiameter
Velocity actual
within ±10% commissioning band
→ Revit · Pipe.FlowVelocity (calc param)
Governing
velocity ≤ 1.2 m/s · BS 5970 insulation per loop temp
Components
Calculation engines
Thermal modelling, Part L compliance (SAP / SBEM), CIBSE TM52 / TM59 overheating, ventilation rates, water demand, electrical load schedules, voltage drop, and earth-fault loop impedance. Inputs versioned, outputs reproducible.
AutoCAD ↔ Revit bridge
Bidirectional synchronisation of geometry, metadata, and parametric relationships. 424 production C# files. Engineers continue working in their preferred tool; the bridge handles reconciliation in the background.
Specification generators
Rule-based clause assembly anchored to BS EN, CIBSE, and UK Building Regulations. Output classified to Uniclass and NRM2 with citable reference for every assertion.
Schedule automation
Plant schedules, valve schedules, distribution boards, lighting schedules, and drainage schedules generated from the live model. Changes to a fixture propagate everywhere it appears.
Reporting pipeline
EPC inputs, condition reports, RIBA Stage 2 loading calculations, and tender summaries produced from the same canonical data store. No manual transposition between deliverables.
Audit & provenance
Every output traceable to its source assumption, calculation method, and standard reference. Versioned snapshots so any historical issue can be reproduced and defended.
Architecture
Surface
WPF UI
Editor add-ins
Diff viewer
Domain
Element mapping
Geometry transforms
Property reconciliation
Calculation kernels
Adapters
AutoCAD .NET API
Revit API
IFC reader/writer
Excel / Word / PDF emitters
Persistence
Local SQLite cache
Versioned snapshots
Provenance log
File index
MEP Specification Builder
A typed standards database paired with deterministic, tool-bound language models. Every clause selected by rule, anchored to a real reference, and validated against configured BS EN, CIBSE, and UK Building Regulations references.
01 · Standards database
Typed clause graph
Every BS EN, CIBSE TM, IET regulation and Approved Document parsed once into a structured graph. Clauses are typed by discipline, classified by Uniclass, costed by NRM2, and linked to the equipment schedules that will satisfy them.
02 · Deterministic generation
Tool-bound language models
Generation runs at temperature 0 with a fixed seed. The model cannot write a clause; it can only call typed tools that return clauses retrieved from the database. Every numeric value, threshold, and scheme name is grounded in a citable source.
03 · Compliance gate
Validate before save
Every output is re-parsed and checked against rule sets: cross-reference integrity, mandatory citations, scheme conflicts, and discipline coverage. A failed validation rolls the document back. The user never sees an unsigned clause.
CL-Y50-MEC-001.docx
Generating
Smoke ventilation
References
Live mock. Every clause is anchored to a citable reference: no hallucinated standards, no made-up clause numbers. Switch tabs to change discipline.
28 MCP tools across 11 disciplines
Disciplines covered
Heating · Cooling · Ventilation · Hot & cold water · Drainage · Gas · LV electrical · HV electrical · Lighting · Data · Fire detection
Adelphos AI
A construction-design platform built on the same automation principles I delivered at JPA. Apps, agentic services, and managed automations across the RIBA stages — Revit Copilot, Clash Solver, COBie Manager, Specbuilder, Schedule Builder, Document Controller, Report Builder.
Adelphos is the construction-design platform I co-founded with two partners, built on the same automation principles I delivered inside JPA M&E Consultants Group. The platform covers several construction-design workflows: a flagship Revit Copilot, a clash solver, a COBie manager, a deterministic specification writer, a schedule builder, a document controller, a report builder, and a verified calculation-engine layer. Beta release scheduled for Q2 2026.
Every output traces back to a source: the model, a calculation, a decision, a drawing. The agent reads from and writes to the tools the practice already uses. No migration, no quiet retention, infrastructure the customer owns. The target is traceable output that a senior engineer can review, amend, and sign off.
drag · scroll · click
For demonstrative purposes only.
Apps · the agent stack
Flagship · Revit add-in
Revit Copilot
532+ Revit commands and 395 MCP tools exposed to a chat-driven agent. Designs systems with you, runs the model, builds your own commands and plugs in your own tools.
Resolves clashes, with permission
Clash Solver
Connects to your ACC project, dedupes against the previous run, separates real clashes from ghost mismatches, and proposes resolutions. Auditable Revit transactions for every accepted change.
A live COBie deliverable
COBie Manager
Populates 20+ COBie sheets from the live model, validates against BS 1192-4 and the project EIR, and traces every cell back to its source Revit element. Reduces manual COBie collation before handover.
Deterministic specifications
Specbuilder
Reads the brief, the live model and the drawings, then writes a stage-appropriate MEP specification where every clause traces to a decision, calculation, or drawing. NBS, CIBSE, ASHRAE, or your house template.
Schedules from the live model
Schedule Builder
Door, window, room, and equipment schedules pulled from the model with manufacturer data layered in. Sized, selected, and exported. The schedule is always the model, never a stale copy.
Transmittals on demand
Document Controller
Drawing-number matching, sub-1-minute transmittal generation, callable from Teams, email, or phone. Revision tracking and registers run from the model.
Eight reports, traceable claims
Report Builder
Site reports, MEP strategies, option appraisals, IESVE energy reports, clash reports, Stage 2 and 3 submissions, and survey writeups. Reads the model, simulation outputs, calculations, meeting minutes, and site photos.
Verified modules the agent calls
Calculation Engines
SAP, EPC, SBEM, cable, lighting, thermal, pipe and duct sizing. Each engine runs server-side with reproducible inputs and auditable outputs.
Agentic services · managed
AI Finances
Automated reconciliation, rolling 12-month cash-flow forecasting tied to programme milestones, monthly management accounts from live data.
AI Project Management
Programme tracking, risk registers, reporting. Tracks tasks, milestones, risks, and reporting exceptions.
AI Document Controller
Managed document control: transmittals, registers, revision tracking. The agent runs your DC function rather than your team running it.
Email services · one-off work
COBie Generator
Forward a model, receive a validated COBie deliverable.
Revit Modelling Agent
Email a markup, receive the Revit model update.
Schematic Designer
Describe the system, receive a production-ready schematic.
Specification Writer
Forward a brief, receive a complete MEP specification section.
One email address per service. Send the brief or model, receive the deliverable. Repeatable template process, dedicated to your project, no data stored for training.
Integrations · works in your tools
Co-founder responsibilities
Platform architecture
Module boundaries, data model, and API contracts across Revit Copilot, Clash Solver, COBie Manager, Specbuilder, Schedule Builder, Document Controller, and Report Builder. Calculation-engine layer separated as a callable service.
AI & automation strategy
Where deterministic logic ends and generative components earn a role. Tool-bound agents, structured-output validation, evaluation harnesses, and prompt-injection defence on regulated workflows. Determinism-first by default.
Engineering hiring & roadmap
Technical hiring across software, AI, and surveying. Sequencing the build to keep engineering load realistic alongside early commercial commitments. Trade-off calls between depth and breadth.
Market & commercial
Segment selection, pricing, channel partner assessment, and the route to first commercial deployments. Investor and board-level commercial decisions.
Customer development
Practitioner interviews across MEP, structural, and architectural disciplines. Mapping where existing tools fail and what a senior engineer can review, amend, and sign off.
Operating principles
Deterministic first
Generative components are used only where they earn their place. Anything touching a regulated deliverable is reproducible, auditable, and reference-cited.
Lives in your tools
Reads from and writes to Revit, AutoCAD, IESVE, Word, Excel, and Trimble. No migration, no parallel platform — the agent works inside the files you already keep.
No data retention
Uploads live on AWS or Azure infrastructure the customer controls. No model training on customer data, no quiet retention.
Audit trail by default
Every agent action — clash resolution, schedule edit, transmittal — records who, what, why, and from which source. The audit trail is the deliverable, not an afterthought.
Live at adelphos.ai · partnership with Jordan Jones and Jean-Luc Wright · beta release Q2 2026 · customer-owned AWS / Azure infrastructure · no data stored for training · the agent reads and writes the same Revit, AutoCAD, IESVE, Word, and Excel files the practice already uses.
WYRMData
Plan, price, and procure AI infrastructure in one workspace. Workload in, rack design out — then a live Revit MEP model generated against manufacturer clearances, auto-routed pipework, and clash detection on the output. AutoCAD export for non-BIM consultancies.
Alpha testing · data.wyrm.ai
From workload spec to installed rack — one workspace. The same deterministic spine then extends into a live Revit MEP model: typed families, parameter-driven plant, auto-routed pipework, and clash detection run on the output. AutoCAD export for the consultancies still working in 2D.
WYRM Data is the planning and procurement layer for AI infrastructure. A buyer types the workload — model size, trainee count, region, budget — and the rule engine returns an auditable rack design: GPUs and chassis to the SKU, cooling strategy by density, kW budget, full BOM. A 3D viewer shows the build before the operator signs anything. A filtered shortlist of UK and EU colos comes back ranked by power density, Tier rating, and certification fit. RFQs go out in one click and replies are parsed into a comparable line-item view — not a PDF stack to chase.
Where the platform sharpens is the Enterprise+ layer, where the rack design becomes a live Revit model. The same rule discipline that picks the hardware drives the families, sizes the plant, and routes the chilled water, condensate, fire suppression, and power containment around every obstacle in the room — manufacturer clearance envelopes honoured, hot/cold aisle topology preserved, clash detection run against the output before the file leaves the platform. The Revit interface keeps the MEP team in the tool they already use; a 2D AutoCAD export covers consultancies that haven’t made the BIM transition.
60+
UK + EU colos
£19/mo
Pro tier
£49/mo
Enterprise base
Usage
Enterprise+ modules
Stack: Next.js 15 · FastAPI · Supabase · Three.js · Resend inbound (RFQ replies via plus-address routing) · LemonSqueezy (MoR). Enterprise+ modules metered per call. Sibling product line to WYRM Procure / Legal / Cyber under the same agentic spine.
Roadmap · where WYRM Data is going
In a series of phased releases, the goal of WYRM Data is to produce an advanced Stage 3 MEP design package. To align with the RIBA and BSRIA BG6 requirements, WYRM will assist in the production of Revit models — not just block placeholders — alongside specifications, schedules, TM54, TM65, BRUKL, and the other items essential for submission. The aim is to reduce the stress and workload on engineers by up to 90%.
Design-stage deliverables already in the platform include:EPCs · NABERS UK · BRUKL · PV (sun-path + shading) · TM54 · TM65 — each one a printable, audit-ready PDF the operator can hand to an accredited assessor as the design-stage briefing pack. Two of them are playable below.
Interactive · 3D data centre
click anything · cabinets · switches · CRAC · raised floor
A scale model of a working cold-aisle: compute and storage cabinets, top-of-rack switches, overhead fibre tray with live traffic, CRAC unit with return-air path, and the raised-floor grilles feeding chilled air up into the racks. Click any element to inspect its sizing, throughput, thermal envelope, and the standards it’s bound to. The same scene is what a buyer sees inside WYRM Data when reviewing a recommended build.
For demonstrative purposes only.
Interactive · NABERS UK indicative rating
BRE NABERS UK · design-stage variant
A scaled-down replica of the NABERS UK module from the WYRM Data design dashboard. IT load and PUE come in from Sizing; on-site PV comes in from the PV stage; the operator declares the tariff mix here. Stars and drivers update live so a buyer can play with the envelope before commissioning an accredited assessor for the twelve-month audited rating.
Inputs
Operating profile
Interactive · PV sun-path & shading
real cast shadows · drag to orbit · scroll to zoom
The full WYRM Data PV rendering slot, ported into the CV with a pre-painted south-facing array on the roof of a three-storey office block. A neighbouring building and a mature tree sit just south of the roof so the morning and evening sun push their shadows up across the modules. Move the day-of-year and hour sliders, or jump to a reference solstice / equinox, and watch the real directional-light shadow walk the array. Shaded modules tint dark and derate at 12% per MCS in the yield calculation behind the scene.
Enterprise+ · MEP design pipeline
The Enterprise+ layer is where the rack design becomes a live Revit model. Two modules carry the weight — the MEP auto-router and the clash-avoidance pass on the output — and the Revit interface keeps the result native to the BIM consultancy that will issue the drawings. A 2D AutoCAD export is provided for projects still working outside BIM.
MEP auto-routing
Chilled water, condensate, fire suppression, and power containment placed as Revit pipework and cable tray, routed around every obstacle in the room model. Hot/cold aisle topology, manufacturer clearance envelopes, and floor-loading constraints honoured before a single segment is drawn. System types and slopes assigned natively — no post-import cleanup.
Clash avoidance on the output
Clash detection runs against the generated route in the live Revit model before the file leaves the platform — pipework vs structure, pipework vs containment, pipework vs the rack envelope itself. The MEP team opens a coordinated model, not a clash report.
Native Revit interface
Round-trip with the colo building model. The output is a real Revit model — typed families, instance parameters, system-classified MEP curves, shared parameters bound to the rack design — opened directly by the engineer, not imported from a PNG or a flat DWG. AutoCAD DWG export available for projects still drafting in 2D.
Revit plantroom populator
Chillers, CRAHs, pumps, valves, and headers placed as Revit families and sized to the rack design’s kW and cooling strategy. Connectors wired, system types assigned, equipment schedules populated. The plantroom is then handed to the auto-router — a single source of truth from rack to plant.
Layout planner with manufacturer clearances
Algorithmic 2D and 3D layout from the same rule discipline as the recommender. Family libraries, service-access envelopes, and aisle widths sized to the recommended rack count — driven straight into the Revit model.
Enterprise+ modules bill per call as metered subscription items on top of the £49/mo base. Routing topology, family libraries, and the specific clearance rule-set are proprietary to the platform.
What the rest of the platform does
Workload in. Spec out.
Free-text or structured intake parsed into an editable spec — workload type, GPU count, memory, networking, redundancy, geography, budget. The model listens at the door; rules decide what gets built.
Algorithmic rack recommender
Pure-function rule engine picks GPUs, server chassis, CPUs, memory, networking (InfiniBand or Ethernet sized to GPU count), cooling (air vs liquid by density), rack count, total kW, and a full BOM. Auditable, versioned, reproducible.
Colo + vendor RFQs that close the loop
60+ verified UK and EU facilities filtered by Tier rating, kW per rack, carriers, and certifications. RFQs go out to colos and OEMs; quotes come back parsed into comparable line items, not PDFs to chase.
3D rack viewer, before the contract is signed
Recommended build slotted into U-positions in a real cabinet. Click a chassis for the BOM, click a GPU for tuning recommendations — driver, container, BIOS, perf flags — drawn from a curated knowledge base.
Demo workflow
The same flow a buyer runs inside WYRM Data, played at demo speed: chat in a workload spec, watch the rule discipline assemble the rack, regions ranked, RFQs return as comparable line items. Replay resets the timeline.
How it works · the loop closes on the fourth step
1. Tell us the workload
Free-text ("fine-tune a 70B model, two trainees, London, under £8k/mo") or the structured form. Parsed into an editable spec the operator can override line-by-line.
2. Get the build
Rule engine returns hardware down to the SKU — GPUs, chassis, CPUs, memory, networking, cooling, rack count, total kW, BOM. Pure functions, no black box.
3. Pick the home
UK and EU facilities ranked by power density, Tier rating, carriers, and certifications (ISO 27001, SOC 2, PCI DSS, ISO 50001). Geographic fit scored against the spec.
4. Run the RFQs
One-click dispatch to colos and hardware vendors. Quotes come back through plus-address inbound, parsed by Haiku into a structured comparison view.
Who WYRM Data is for
MEP and data-centre design consultancies
The compute brief becomes a live Revit model. Algorithmic spec, clearance-aware layout, auto-routed pipework, and clash detection — automating the layout, routing, and clash-checking steps so engineers can review the generated model. The deliverable opens in the same Revit session the team is already running.
AI startup CTOs and founding engineers
First training cluster, first colo move, or stepping down from cloud bills that doubled overnight. Outputs a kW budget, BOM, colo shortlist, and draft RFQs from the workload brief.
Enterprise procurement teams
One RFQ standard across regions and hardware generations. Comparable quotes back, audit trail intact, no spreadsheet archaeology.
Hyperscaler MEP teams
Run the rack recommender across multiple workload profiles in parallel, then push each variant through the Revit plantroom populator and auto-router to compare CAPEX, kW envelope, and floor footprint as native models side by side.
WYRMProcure
Agentic AI for procurement and supply-chain intelligence. Sanctions, carbon, FX, shipping, and supplier verification in parallel.
Procurement teams currently spend days reconciling sanctions screens, carbon disclosures, currency exposure, freight routing, and supplier background checks before a purchase order can be signed. Procure does that work in one pass. A buyer types what they want, the platform runs the diligence in parallel across the relevant data domains, and a ranked shortlist comes back with the customs and compliance paperwork already drafted. Confidence is scored across an ensemble of forecasting models so the recommendation comes with the risk attached, not bolted on later.
The platform runs fully agentically. A supervisor agent decomposes each query into specialist sub-agents (sanctions, carbon, FX, shipping, supplier verification, customs), routes them in parallel, and reconciles the answers through a confidence gate. Each agent monitors its own data sources and self-heals on feed timeouts, schema drifts, and rate-limit failures: it retries against the canonical source, falls back to a cached read, and flags the degraded path on the audit trail. The workflow is designed for automated triage, with escalation paths when confidence or data quality falls below threshold.
Five procurement dimensions
Sanctions
OFSI · OFAC · EU · UN screened in parallel against parent ownership trees.
Carbon
Lifecycle scoring with CBAM-ready customs documentation.
FX exposure
Spread, settlement, and forward-cover suggestions for cross-border payments.
Shipping
Routes with chokepoint risk, tanker queueing, and port congestion telemetry.
Supplier verification
Capacity, ownership, and audit-grade supplier records — not directory listings.
Sample query
agent.proc.bid
> source 50,000 units of medical-grade nitrile gloves
UK
PremiumUnited Kingdom
Local sourcing · shortest lane
CN
WatchChina
Baseline cost · deepest catalogue
VN
Best matchVietnam
Diversification · stable compliance
Global news intelligence
agent.news.synthesise
42 sources · multilingual · Haiku-translated
- shippingReuters·2m ago
Yemen-based group claims fresh Red Sea drone strike on bulker
→ Suez routings reweighted · ETA +6d on 14 active baskets
- fxBloomberg·8m ago
CNY weakens past 7.32 as PBoC cuts reserve requirement
→ CNY-denominated POs flagged · forward cover refreshed
- sanctionsFT·14m ago
EU adds 19 entities to Russia circumvention list, includes UAE intermediaries
→ 237 supplier records re-screened · 3 promoted to manual review
- shippingLloyd’s List·22m ago
Panama Canal slot auctions surge as draft restrictions tighten
→ US east-coast routings repriced · contingency carriers refreshed
- commodityArgus·31m ago
Indonesian nickel export quota tightened ahead of smelter audit
→ Battery-grade nickel basket rerouted · alt-supplier ranks rebuilt
- carbonEU Commission·44m ago
CBAM definitive period guidance: embedded-emissions methodology update
→ Steel + cement carbon scores recalibrated · audit pack regenerated
Live event stream
/api/agents/stream
streamingChokepoint signal
Strait of Hormuz
25.6° N · 56.3° EOil chokepoint · 20% of global crude
Signals tracked
- AIS tanker queuing telemetry
- IRGC tension score (GDELT)
- Brent crude response curve
- Lloyd's war-risk premium
50+
API routes
20+
Live data feeds
10
Domain agents
23
Dashboards
Stack: Next.js 15 · FastAPI · Supabase · Redis · Railway · Vercel. CesiumJS 3D globe with SGP4 satellite tracking, MapLibre 2D, AISStream maritime WebSocket. Single-developer production build.
Opportunity radar · live tender feed
Contracts Finder · FTS · TED · SAM.gov · PCS · Sell2Wales · CE RSS
Public and private procurement, construction, and manufacturing opportunities ingested from seven government and industry sources. Every row is scored against the buyer profile, then expanded into a WYRM-generated analysis card: direct-purchase savings against the prime, resale margin, and a draft bill of materials. Pro sees a 25-row preview; Enterprise unlocks Bid Assist — a full tender builder with playbook-bound clause generation.
CF-2026-0418 · Ministry of Justice · South-east England
M&E maintenance framework — south-east estate
£12.4M / 4 yrs
closes 14d
50710000 · Building services
Contracts Finder
92
strong fit
14.2%
save £186k
FTS-2026-1027 · NHS Shared Business Services · UK-wide
Nitrile examination gloves — national framework call-off
£4.8M / 18 mo
closes 9d
33141420 · Medical gloves
Find a Tender
88
strong fit
11.6%
save £412k
TED-2026-0772 · Federal Network Agency (DE) · North Sea
Offshore wind cable-laying support · 2026–28
€18.2M / 24 mo
closes 21d
45231400 · Cable laying
TED
71
screen
9.4%
save €340k
PCS-2026-0339 · NHS Greater Glasgow & Clyde · Scotland
BEMS upgrade · acute hospital portfolio
£3.6M / 18 mo
closes 17d
45331000 · Heating ventilation
PCS · Scotland
86
strong fit
13.1%
save £82k
SAM-2026-A219 · US General Services Administration · United States
LED retrofit · federal building portfolio NCR
$5.4M / 24 mo
closes 28d
31518600 · Lighting
SAM.gov
64
screen
8.8%
save $140k
agent.bid.assist · Enterprise
Tender builder
Selecting an opportunity opens Bid Assist. The agent reads the buyer brief, pulls the team tender profile, and drafts every required section against the pinned playbook — method statement, pricing schedule, compliance pack, social value, technical approach. Each clause is footnoted with the playbook section it answers to. The operator edits in place, accepts or rejects per-clause, and exports a paginated PDF + Word pack ready for submission.
- Tender-profile round-trips with WYRM Legal playbook
- Buyer-question handoff to TenderLead drafter
- Buyer-enquiry email drafted + copied to clipboard
- Audit trail of accepted vs rejected clauses
CF-2026-0418 · MoJ M&E framework
Bid Assist · drafting01 · Method statement
playbook ✓Mobilisation in two phases against the 18-month framework: phase 1 covers the four south-east acute sites under direct supervision of a chartered building services engineer (MCIBSE). Phase 2 extends to the eight community sites once asset registers are reconciled. Each site receives a 30-day audit window, a written SFG20 alignment, and a digital twin export to the MoJ FM platform.
02 · Pricing schedule
playbook ✓Fixed-price annual sum of £2.94M Year 1, indexed to BCIS at +1.2% real over Years 2–4. Reactive call-out rates aligned to the team rate card (£68/hr in-hours, £92/hr out-of-hours). Materials at trade list less the framework discount on Schedule B. All sums exclude VAT.
03 · Compliance pack
playbook ✓ISO 9001:2015, ISO 14001:2015, ISO 45001:2018, and SafeContractor certification attached. Insurance: PI £10M, PL £10M, EL £10M. Modern Slavery Act statement, GDPR DPIA, and the team Cyber Essentials Plus certificate enclosed. SFG20 maintenance regime aligned to the asset list at Schedule A.
17 of 17 sections drafted · audit trail attached
Lifecycle carbon · per unit
agent.carbon.score
UK
CBAM-readyUnited Kingdom
CN
CBAM exposureChina
VN
CBAM-exemptVietnam
WYRMLegal
Reviews vendor contracts against a pinned red-line playbook rather than generic UK defaults, and surfaces clauses for legal review. WYRM Legal sits between the vendor draft and the lawyer.
Every draft passes through a tool-bound agent stack — clause extraction, party detection, liability cap and indemnity scope scoring, governing-law check, termination-notice match, GDPR and Standard Contractual Clauses lookup, and a final playbook alignment score. Each finding is cited back to the pinned playbook section it answers to, so the lawyer reading the verdict sees the rule, the deviation, and the exact reference.
While documents are being generated, the dashboard shows an interactive loading state.
Legal Pro
£19/mo
10 reviews / mo
Legal Enterprise
£49/mo
unlimited · SSO · API
01 · Playbook engine
Your red-lines, pinned once
Enterprise users author a commercial playbook at /settings/legal-profile: liability cap, indemnity scope, payment terms, termination notice, governing law. Every review and draft inherits it. The same playbook drives Procure bid-assist contract review.
02 · Deterministic generation
Eleven document types
Services agreement, NDA, Framework, DPA, SOW, LOI, employment offer, software licence, supply, reseller, settlement. Each runs a calibrated system prompt and is tool-bound to clause lookup, citation, and playbook scoring. No invented case law, no hallucinated references.
03 · Word and PDF export
Tracked-change document export
A second Sonnet pass absorbs the verdict recommendations into the contract body. The rewritten document emits as a clean .docx with tracked-change comments and a paginated .pdf. Cost is around £0.04 to £0.15 per export pass; format toggles are free.
acme-msa-v2.1.docx
Drafting
Master services agreement
Acme Limited (Customer) ↔ Wyrm Holdings Ltd (Supplier)
Playbook citations
Export with absorbed recommendations · second Sonnet pass
Live mock. Every clause cites the pinned playbook section it answers to. Switch tabs to flip between the five document types WYRM Legal generates and reviews.
22 MCP tools across review and drafting
Document types covered
Services agreement · NDA · Framework · DPA · SOW · Letter of intent · Employment offer · Software licence · Supply agreement · Reseller · Settlement
WYRMHealthcare
A vendor-neutral go-live and operational-memory layer for NHS providers. WYRM Compass sits inside Teams during EPR / PAS launches, citing the trust's own SOPs at the moment of need; WYRM Migrate ports records out via GP Connect; WYRM Manage compounds the assistant into permanent institutional memory.
Every major NHS digital programme — EPR replacement, PAS upgrade, ICS shared-care-record stand-up — fails at the same point: the moment systems flip live and clinicians can’t find what the SOP says to do. The big consultancies charge £10–50m to programme-manage that risk and still produce disasters. WYRM Healthcare ships an operational answer instead: a vendor-neutral go-live assistant that sits inside Teams during launch, cites the trust’s own SOPs at the moment of need, and compounds into the trust’s permanent institutional memory once the dust settles.
The trinity composes commercially. Compass lands the relationship during go-live. Migrate is sold during contract renewal or ICB consolidation. Manage compounds into multi-year ARR. Any single pillar is a viable standalone business; all three together is a category. Critically, the trinity does not require winning the EPR substrate sale — Healthcare is commercially viable as a vendor-neutral go-live + migration + management business without it.
3
Product lines in the trinity
4w
Phase H1 demo plan
18
Versioned SOPs (Ashmere Vale)
Haiku 4.5
Locked model
Stack: Next.js 15 · FastAPI · forked Medplum (Apache-2.0, FHIR R4) behind a WYRM-owned FHIR UK Core gateway · pgvector + voyage-3 retrieval · Claude Haiku 4.5 under a no-retention contract · GP Connect Access Record: Structured for migration. Fifth sibling product line under the same agentic spine as Procure / Cyber / Legal / Data.
The trinity · how the three lines compose
WYRM Compass · go-live + operations
A vendor-neutral, clinical-safety-cased AI assistant trained on the trust's own SOPs, runbooks, escalation matrices, and architecture docs. Sits inside Teams during EPR / PAS / shared-care-record launches and answers the question every clinician has at 02:00 — "what does our SOP say to do right now?". Works on Epic, Cerner, EMIS, TPP, Oracle Health, or the WYRM-owned substrate.
WYRM Migrate · GP Connect ARS replay
Patient-record migration that does not require incumbent vendor cooperation. Uses the NHS-mandated GP Connect Access Record: Structured channel to stream every record into a normalised UK-Core FHIR projection, then replay into the destination EPR on a per-patient idempotent contract.
WYRM Manage · permanent institutional memory
Compass ages into the trust's long-term operational layer. Booking, NHS Notify, QOF dashboards, ShCR contributions, and analytics surfaces compound on the same agentic spine. The go-live assistant earns its way into multi-year ARR by becoming the place the trust's knowledge lives.
Interactive · See the trinity in action
wyrm.ai/healthcare#demo
The same tab-switched widget that ships on wyrm.ai/healthcare. Three working demos for the three pillars — Compass takes typed questions and runs a real lookup over the trust's SOP corpus, Migrate runs a dry-run pipeline over a chosen source EPR + cohort, Manage drives a live booking optimiser and operational-memory surface. All data fictional; production deployments are gated by a per-trust DCB0129 Safety Case.
Tap a preset or type a procedural question — try one about an allergy or downtime SOP.
Interactive demonstration. All trust documents, EPR data, and operational metrics shown are fictional. Production deployments are gated by a per-trust DCB0129 Safety Case.
Interactive · Clinical safety + compliance
DCB0129 · DSPT v8 · DTAC v2 · WORM audit
Anyone can spin up a RAG over a trust's SOPs in a week. The clinical-safety substrate — DCB0129 Safety Case, DSPT v8 with CAF-aligned independent audit, DTAC v2, a named Clinical Safety Officer on retainer — takes twelve months and is gate-kept by senior clinician availability. WYRM Healthcare carries that posture so individual NHS deployments don't have to.
DCB0129 Safety Case
Per-trust safety case with named hazards, controls, and a hazard log maintained across deployment versions. Reviewed by your CSO and ours every quarter.
DSPT v8 + CAF-aligned audit
Standards-met with the 2025/26 CAF-aligned independent-audit requirement that now applies to in-scope IT suppliers.
DTAC v2 pack
Full DTAC v2 (mandatory from 6 April 2026) coverage across clinical safety, data protection, technical security, interoperability, and usability.
Named Clinical Safety Officer
A senior clinician on retainer with current GMC or NMC registration and clinical-safety training. Available to your CSO for joint safety reviews.
Cyber Essentials Plus
Independently audited annually. UK-only data residency on AWS London or Azure UK South.
Immutable audit log
Every Q/A persisted to an append-only WORM store with 25-year retention. Every refusal carries reason, escalation path, model version, and audit ID. Surfaced for incident review without ceremony.
- ·PAT-001 v2.3 (rank 1, score 0.92)
- ·PAT-001 v2.2 (superseded, rank 2, 0.78)
- ·DOWN-001 v1.7 (rank 3, 0.41)
Returned PAT-001 §3.1 v2.3 procedure with PDS-trace cross-reference. 178 tokens · 412 ms.
Pre-pilot — no clinical data, no NHS integrations live. The demo carries a DCB0129-mini safety statement and is explicit that it is a demonstration of the safety pattern, not a clinically-safe deployment. Production deployments are gated by a per-trust DCB0129 Safety Case.
WYRMCyber
Sixteen cyber tools under one subscription. Threat intel, OSINT, CVE monitoring, dark web, recon, attack surface, sanctions, and a forensics data toolkit. Built on the same agent and intelligence stack as Procure, repointed at cyber risk.
Cyber spawned from the same intelligence platform as Procure. The agents that screen suppliers and reconcile shipping for procurement repoint at IOC feeds, CVE pipelines, dark-web crawls, and OSINT tools when running cyber workloads. Combines recon, OSINT, CVE monitoring, dark-web search, sanctions screening, and forensic data tools in one interface.
Every tool is wrapped in an agent. A single brief — a domain, a handle, a CVE, an IP — fans out across the 16 tools in parallel, correlates the findings, and returns one ranked verdict instead of 16 raw reports. The triage agent auto-diagnoses by chaining recon → fingerprint → CVE → threat intel → dark web → sanctions, weighing severity against EPSS exploit probability and KEV status. Watchlists run continuously and only surface when something material changes — no analyst is paid to re-read clean reports. Degraded feeds are flagged honestly rather than silently dropped, so the verdict is always traceable to the evidence it stands on.
Pro
£19/mo
Enterprise
£49/mo
Live recon console · sample workflow
one brief · six tools chained · one verdict
wyrm-cyber · brief: due-diligence on a Tier 2 supplier
$ recon --target acme-supply.co.uk --all
Risk score
62/100
Engage with caveats · KEV CVEs + C2 match require remediation evidence before signing.
Identity & OSINT
Find people, profiles, and the trail they leave
Correlate usernames, phone numbers, email addresses, and image metadata across OSINT sources.
Username Search
Map a handle to every social, forum, and dev account it lives on. 3,000+ sites in a single sweep.
Phone OSINT
Validate a number, identify carrier and region, and surface the public footprint tied to it.
Email Breach
Check an address against known breaches — passwords, sessions, PII exposure dates, breach source.
Image OSINT
Reverse-image search, EXIF metadata extraction, and provenance hints. See where a photo lived and when it was shot.
Threat Intelligence
Know what's targeting your sector before it lands
Curated IOC feeds, real-world CVE triage, and entity sanctions in one screen.
Threat Intel
Aggregated feeds from CISA KEV, abuse.ch, AlienVault and 20+ sources — scored, deduped, correlated.
CVE Monitor
Live NVD + EPSS exploit-probability + CISA KEV. Triage actual risk, not raw CVSS. Watchlists by vendor or CPE.
Threat Feeds
Curated IOC streams — malware hashes, C2 infrastructure, phishing kits, ransomware claims. Webhook-ready.
Sanctions
OFAC, UK HMT, EU consolidated, OpenSanctions. Screen entities against every major regime in one query.
Recon & Attack Surface
External attack-surface mapping and recon tooling
Subdomain discovery, port scans, light stress probing, dorking, and historical snapshots.
Recon
Subdomain enumeration, certificate transparency, tech fingerprinting. Full external footprint in seconds.
Network Scanner
Port scans, service detection, banner grabs, and light stress probing. See what's open and how it holds up under load.
Attack Surface
Continuous monitoring of exposed assets, scored by criticality. Catches new shadow IT before attackers do.
Google Dorking
Pre-built advanced operators for finding exposed docs, configs, panels, and forgotten subdomains.
Dark Web & Forensics
Search the corners search engines won't reach
Onion crawlers, paste-site mentions, IP attribution, and a full data-transformation toolkit.
Dark Web
Onion sites, paste sites, ransomware leak forums. Search for your domain, staff, or clients in the underground.
IP Lookup
ASN, geolocation, hosting reputation, abuse history. Residential proxy? VPN exit? Datacenter? Known C2?
Data Toolkit
Decode, transform, extract, hash, encode, decrypt. The CyberChef-style swiss-army knife for evidence work.
Wayback
Historical site snapshots. See what was live, when it changed, and what got quietly removed.
Premises Presence · Enterprise
See the room — without watching anyone in it
Camera-free, wearable-free occupancy intelligence. A premises' Wi-Fi sensing mesh detects how people move through the building from the way bodies disturb radio signals; WYRM turns that stream into anonymous, alertable presence intelligence. The visualisation below is the full demo — drag a figure to pick it up, hover a room for live occupancy stats, and watch the alert fire on restricted-zone entry.
Drag a figure to pick it up · hover a room for occupancy stats · toggle Health Watch to preview the Phase 3 medical-alert demo. Illustrative simulation — not a live feed. People settle at desks and pause to talk, but only anonymous skeletal pose (arm and hand kinematics, gait, gesture) is reconstructed from Wi-Fi signal disturbance — never camera images, never audio, and never identity. No cameras, no wearables. Health Watch is opt-in roadmap (Phase 3) with per-subject consent — it is not enabled today.
Why Wi-Fi sensing
Cameras vs Wi-Fi sensing
Two ways to know a room is occupied. Only one of them is private by default.
The old way
Cameras & Wearables
- Captures faces and identity — every frame is personal data
- GDPR / HIPAA exposure by default; DPIAs, retention and access controls on every feed
- Dead zones behind walls, partitions and furniture
- Fails in smoke, dust, glare and darkness
- Wearables assume willing, charged, correctly-worn users
The WYRM approach
Radio, not pixels
- Radio signals only — no images are ever captured
- Physics-based: reads Fresnel-zone and multipath disturbance, not pixels
- Works through walls, furniture, smoke and total darkness
- Zero user compliance — ambient by design, nothing to wear or charge
- Privacy-preserving and consent-gated from the first byte
The roadmap
Three phases, each separately consented
The capability ladder is additive — and that's the point. Each phase is enabled deliberately, with its own consent. Nothing escalates silently from a headcount to a heartbeat.
Phase 1
Presence & Occupancy
Anonymous headcount per zone with roster-baseline anomaly detection. Identify additional personnel in the building without identifying anyone in it.
- Anonymous headcount per zone
- Roster-baseline anomaly detection
- After-hours presence
- Unexpected occupancy
- Headcount-over-roster
- Restricted-zone access
- Sensor-offline detection
Phase 2
Activity
Movement and activity patterns — dwell time and flow between zones. No identity from the signal itself; behaviour in aggregate, never a named individual.
- Movement and activity patterns
- Dwell-time analysis per zone
- Flow between zones
- Optional employee identification via consented badge / access tags — never from the radio signal itself
Phase 3
Health module
A wellbeing feature, separately consented. Coarse vital signs from Wi-Fi CSI for lone-worker safety and fall / distress detection — opt-in by deployment, never on by default.
- Vital signs (heart rate, respiration) via Wi-Fi CSI
- Lone-worker safety monitoring
- Fall and distress detection
- Opt-in / separately consented — wellbeing, not surveillance
Built for
Pen testers
Recon, attack surface, network probing, and dorking in one console — without firing up six VMs.
Blue teams · SOC
Triage threat intel, screen IOCs against your stack, and pivot from a single alert into the full picture.
OSINT analysts
Username, image, dark-web, and breach data — track a story from a fragment to the source.
DFIR · bug-bounty
Wayback snapshots, EXIF chains, IP attribution, and a data toolkit for evidence-grade analysis.
16
Cyber tools
8
Live RSS feeds
4
Sanctions regimes
20+
IOC sources
Live at cyber.wyrm.ai · NVD + EPSS + CISA KEV CVE pipeline · OFAC, UK HMT, EU consolidated, OpenSanctions screening · 8 direct-RSS news feeds · authFetch pattern across all tool pages · degraded-state surfacing for partial feed failures.
PropTurn
A property-management platform covering sales, lettings, supported living, and pre-purchase due diligence under one tenant. Auto-orders conveyancing searches, pulls EPCs from gov.uk, and runs the deal analyser, consolidating work that usually spans the Guesty / PriceLabs / Turno / Notion stack.
PropTurn is a property-management platform for the UK residential, short-let, and supported-living markets. One platform consolidates the channel-sync, pricing, cleaning, and communications work that usually spans the Guesty / PriceLabs / Turno / Notion stack — and extends past it into the work that usually lives in a separate conveyancing or analyst tool. Pre-purchase searches are auto-ordered through tmgroup on instruction; EPCs pull from the gov.uk Open Data API; the deal analyser scores yield, refurb spend, and exit value before the offer is signed. Lettings ops carry tenancies through payment scheduling, comms, and maintenance dispatch on the same spine. Landlords sign up to a 15% full-service model; property managers bring their existing portfolio onto a single tenant. A PM Copilot reads contractor proposals, sense-checks them against the team rate card and trade playbook, and proposes value-engineering moves before the job is awarded.
UK residential · short-let · supported living
One operating system for the property, the portfolio, and the people servicing both.
Four audiences · one platform
Residential · short-let · supported living
Landlords
Full-service letting and management on a 15% revenue share. Compliance, dynamic pricing, dispatch, settlement.
Multi-portfolio, multi-region, single tenant model
Property managers
Replaces the Guesty / PriceLabs / Turno / Notion stack with one operating system. Channel sync, pricing, cleaning, comms.
First job free · monthly subscription thereafter
Service partners
Cleaners, linen suppliers, maintenance trades and gardeners receive jobs through a marketplace with SLA tracking.
Sales · lettings · cross-sell to the same vendor
Estate agents
Sales progression, conveyancing search ordering, EPC pulls, and deal analyser pre-offer — sitting alongside the lettings book on the same tenant.
Operator console · live simulator
- Boiler · NE1/22SLA 2h
- Lockout · NE2/4keys · 14m
- Low stock · NE3/8re-order
- New booking · NE4/5turn +1d
- Payouts cleared£18,420
- Owner statements14 drafting
- HMRC VAT Q2£12,640 accrued
- Refunds pending£215
Pre-purchase pack · EPC + utility searches
On instruction PropTurn auto-orders the conveyancing search bundle from tmgroup and pulls the property’s EPC from the gov.uk Open Data API in parallel. Each response is normalised server-side — utility connection sizes, flood class, sewer adoption, and contaminated-land linkage — and the file surfaces only the facts that move the deal. The full responses stay in the case file for the solicitor.
Pre-purchase pack
auto-assembled · 4 of 5 returned14 Elm Grove, Wilmslow, SK9 3PR
Energy Performance Certificate
gov.uk · pulled 09:14
Current
Potential
Expires
21 Mar 2030
Surfaced recommendations
- Internal wall insulation£170/yr · C 68 → C 74
- Replace boiler with air-source heat pump£240/yr · C 74 → B 81
- Solar PV 3.5 kWp£420/yr · B 81 → B 88
Conveyancing searches
tmgroup · ordered 14:02
Energy & utility connections
returnedAll four primary services connected. Fibre available, copper retired 2025.
- Mains gas · Cadent · 25 mm MDPE · 4.1 m from boundary
- Electricity · NGED · 32 kVA available · single phase
- Water · United Utilities · 25 mm MDPE · 3 m from boundary
- Sewerage · public foul + surface · adopted 6.8 m rear
- Fibre · Openreach FTTP · 900 Mbps · 2-day install SLA
Drainage & water (CON29DW)
returnedPublic foul + surface water. Adopted sewer 4.2 m from rear boundary — no building over.
- No build-over agreement required
- Annual charge basis · metered
- No drainage in adopted highway
Environmental (Groundsure Avista)
returnedLow flood risk. Former mill 220 m W — no contaminated land linkage.
- Flood · 1 in 1,000 surface · 1 in 1,000 river
- Radon · class 1-3% · basic measures only
- Subsidence · low · no historic claims in postcode
Local authority (LLC1 + CON29)
returnedAdopted highway. No s.106. Listed s.215 notice from 2021 — closed.
- No outstanding enforcement
- Conservation area · Wilmslow Town
- No planning constraints recorded
Planning & building history
pendingOrdered 14:02 today · LPA portal SLA 48 h.
Pack assembled from the gov.uk EPC API and tmgroup conveyancing feed. PropTurn auto-orders the matching search bundle on instruction, normalises every response, and surfaces only the facts that move the deal — the rest sits in the file for the solicitor.
PM Copilot · contractor proposal review
PM Copilot
contractor reviewHansen Electrical quoted £4,820 for the rewire at Flat 3, Maple Court. Three-day job, includes Part P. Sense check before I sign off?
Two flags on this one. Your rate card sits at £3,650 mean across 11 historical electrical jobs of similar scope — Hansen comes in roughly 32% over. The bigger issue is labour: their day rate works back to £62/hr against a playbook ceiling of £55/hr. On the materials line, specifying 6242Y twin-and-earth in standard 1.5 and 2.5 mm² sizes instead of LSF cable would normally drop materials around 18% without changing Part P compliance. Worth asking for a revised quote on that basis. For definitive sign-off, run the revised number past your retained electrician or NICEIC body.
Can we get the EICR testing and certification included rather than billed separately?
Cited against the team rate card and playbook. The Copilot will not invent £ amounts, percentages, or regulatory references — anything it quotes is grounded in attached context. For definitive sign-off on tax, contract law, or structural matters it defers to your accountant, solicitor, or surveyor.
47
Platform integrations
36
Operational categories
1,209
Test cases
15%
Management fee
Operational coverage
Stack: Next.js 15 · NextAuth · Prisma 6 · DaisyUI · Supabase · Railway Redis · Stripe · @ai-sdk/anthropic. Channel sync against Booking.com, Airbnb, Vrbo, and direct booking. Pricing agent runs against a 14-day forward demand model. Conveyancing searches via tmgroup; EPCs via the gov.uk Open Data API; deal analyser scores yield, refurb spend, and exit value pre-offer. 1,209 tests across 21 modules.
Plan-MI
A trip and event planning prototype. Generates itineraries, bookings, group coordination, multi-currency settlement, language learning, and wedding planning from a single brief.
Plan-Mi is a trip and event planning prototype. One brief produces an itinerary, the bookings to back it, the language the traveller needs at each stop, a group ledger to track shared finances, and a wedding-planning workspace for trips that happen to also be weddings.
One brief · five workspaces · zero prompts to refine
Autonomous itinerary
A single trip brief produces flights, transfers, accommodation, restaurants, and reservations.
Language tutor
Pre-trip vocabulary aligned to the cities visited. Restaurant phrases, transport, customs, and pronunciation drilled in 5-minute sessions.
Wedding planner
Venue, guest list, RSVPs, vendors, schedule, dress codes, gift registry, and travel logistics for the wedding party in one workspace.
Group settlement
Multi-currency expense tracking with FX-aware splits. The ledger settles to one transfer per person at the end of the trip.
Three-day itinerary · Lisbon → Sintra → Évora
Lisbon
TueCoffee at Fábrica. Walk Alfama. 19:30 dinner at Cervejaria Ramiro · booked.
Sintra
WedQuinta da Regaleira 10:00 · skip-the-line tickets. Cascais train 16:14.
Évora
ThuComboios CP 09:21 from Oriente. Templo Romano. Dinner Fialho 20:00 · booked.
Wedding workspace
Wedding · Florence
41 days out14 June · Castello di Vincigliata, Florence
Guest list
94 confirmed · 12 awaiting · 4 declined
Vendors
Florist · band · photographer locked
Schedule
Friday rehearsal · Saturday ceremony · Sunday brunch
Travel
38 international guests · transfers booked
Language tutor
Lisbon · Portuguese
Day 4 · 38%Uma mesa para dois, por favor.
A table for two, please.
A conta, por favor.
The bill, please.
Onde fica a estação de comboios?
Where is the train station?
Group settlement
- Jamespaid €612settled
- Annapaid €0owes €186
- Leopaid €0owes €242
- Miapaid €0owes €184
FX rates locked at booking. One transfer per person on return.
0
Prompts to refine
6
Currencies handled
98%
Mobile traffic
<10m
Plan to confirm
Stack: TypeScript · Express · React · Vite · SQLite. Mobile-first by default with 44 px touch targets, full-screen modals, and platform capture handlers. Booking layer wraps flights, rail, hotel, and restaurant providers. Language and settlement modules run client-side.
Agentic Algorithmic Trading Platform — Apollo
A three-arm execution platform: BTC spot and perpetuals running live, a UK ISA arm for equity factor strategies under HMRC stocks-and-shares allowance, and a CFD arm wired into Interactive Brokers for multi-asset leverage. A shared agentic prediction layer feeds all three — deterministic gates carry the signal weight; LLM-mediated agents add regime context the gates cannot derive directly.
Target runtime is an NVIDIA Jetson Orin Nano Super — 8 GB Ampere, ~67 INT8 TOPS, NVMe-booted, active-cooled on UPS — that carries the three execution arms, the Kronos forecasting sidecar, and the operator dashboard in a single £199 device. The agentic council routes to cloud inference or a local 7–13B model on the same hardware; with local routing the recurring cloud LLM spend collapses to near-zero, kept only as a surge fallback. Steady-state power draw is ~15 W (~0.36 kWh/day, well under 15p/day at UK domestic rates) versus the ~$60–120/mo a cloud VM running the same stack 24/7 would carry. The deterministic execution path is unchanged either way; the LLM substrate is an interchangeable component, not a load-bearing dependency.
Target runtime · NVIDIA Jetson Orin Nano Super
Funding (8h)
0.0050%
Binance perps
Fear & Greed
61
Greed
Liquidations 24h
$180M
Long + short
Open interest
$4.60B
24h Δ
Agent ensemble
EMA cross gate
EMA 10 × 20 × 50
deterministic
RSI momentum gate
HA-RSI 14
deterministic
VWAP reversion
VWAP, Volume
deterministic
Funding & OI gate
Perp funding, OI Δ
deterministic
Liquidation pressure
Liq heatmap
deterministic
Market structure
HH / LL, BOS
deterministic
Sentiment synthesis
F&G, social NLP
predictive
Smart-flow predictor
CVD, order book
predictive
Predictive gate
signal firedSHORT
78% aggregate confidence
Leverage
5×
Hold suggested
4h 38m
Weighted by agent track record and live spread / liquidity gates. Output gated again before any execution touches a venue.
Indicator stack
- EMA 10 / 20 / 50 / 200
- VWAP
- Heikin-Ashi RSI
- MACD
- Order-flow imbalance
- Higher-high / lower-low
- Volume profile
- Funding rates
- Liquidation heatmap
- Crypto Fear & Greed
- Sentiment synthesis
- Open-interest delta
Portfolio · paper
$127,765.76
+44.8% · 42 days
Win rate
58.4%
Avg R:R
1.86
Max DD
−8.2%
Sharpe
2.41
Market perception feed
headline → odds shift
- reuters·04m
Fed signals openness to crypto market structure bill ahead of Q1 markup
market · BILL passes by EOY
41%→58%▲17 - bloomberg·11m
BlackRock IBIT logs eighth consecutive day of net inflows
market · BTC > $135k by Jun
46%→53%▲7 - reuters·23m
SEC delays decision on multi-asset spot ETF basket to Q2
market · Approval before May
62%→44%▼18 - wsj·47m
MicroStrategy raises $1.2B convertible to expand BTC treasury
market · BTC closes green this week
51%→59%▲8 - ft·1h
EU MiCA Phase-2 stablecoin caps land softer than industry expected
market · USDC EU listings hold
47%→66%▲19 - coindesk·1h
Tether quietly mints $1B USDT on Tron, on-chain trackers flag
market · BTC > $130k by Fri
38%→49%▲11
Headlines parsed by the news agent, mapped to internal probability markets, and re-priced against the consensus before the next bar prints.
Position sizer
Notional
$84,990
Margin used
$16,998 · 13.3%
Kelly criterion · f* = (bp − q) / b
p
0.78
win prob
q
0.22
loss prob
b
1.86
R:R
f*
66.2%
full Kelly
Apollo runs ½ Kelly · 33.1% of equity per trade. Full Kelly maximises log-growth but lives at the variance edge.
Monte Carlo · 400 paths · 60 trades
p5
×167611.35
bad path
median
×33180488.32
typical
p95
×2720814314.50
good path
Ruin (bal → 0): 0.0% · Halved (bal < 0.5×): 0.0%. Resampled on every consensus update.
Market predictions · ensemble
+0.4%
62% probability
EMA cross + flow
+2.1%
71% probability
Funding flip · liq cluster cleared
+5.6%
58% probability
Trend regime + sentiment
±3%
49% probability
Range-bound macro
Tier-C runtime mathematics · selected
src · Apollo Mathematics Foundations §12
Prediction-interval form: variance grows with distance from the calibration mean. Apollo cannot claim high confidence on a forecast outside its calibration range.
The LCB multiplier k tapers from cold-start 97.5% to steady-state 84% as effective sample size accumulates. Apollo cannot relax confidence by waiting through a long no-trade period.
Maximise expected log-wealth subject to a 5% probability bound on a 252-day max drawdown breaching 25%. Solved by bisection on f with a fixed-seed residual-block bootstrap.
Closed-form schedule that minimises the variance-penalised expected slippage of liquidating Q₀ over T seconds. Shipped in P3; the executor falls back to TWAP when ϕ = 0.
The execution-layer Hamilton–Jacobi–Bellman equation that lets the trade rate respond to an admissible micro-drift signal. Strict superset of Almgren–Chriss; activates only after walk-forward + shadow promotion.
Card 19 (sizing) · Card 04 (Almgren–Chriss) · Card 03 (Cartea–Jaimungal) · Busseti–Ryu–Boyd 2016 · Grinold 1994. One mechanism — state-conditional LCB-Kelly with a drawdown-probability constraint. One knob — k. No LLM in the decision path.
Prediction Market Bots — Kalshi + Polymarket
Two production bots running on regulated and decentralised prediction-market venues. Kalshi BTC/ETH/SOL/XRP fifteen-minute binary options with a SNIPE-first edge plus Safe Compounder passive NO-side scanning; Polymarket multi-strategy arbitrage covering latency arb against Binance spot, Dutch-book risk-free arb, and cross-venue spreads against Kalshi.
Strategies in production
Expiry tail-probability snipes on KXBTC/KXETH/KXSOL/KXXRP 15-minute contracts. Importance sampling tunes the prior, sigmoid-fair probability sets the entry threshold.
Passive NO-side scanner across the entire Kalshi catalogue. Edge-weighted by Brier-calibrated win-rate; Kelly-fractional sized; pulls in slow compounding income while SNIPE waits.
Binance BTC/USDT WebSocket feeds Polymarket Up/Down 15m & 1h re-pricing. Acts on momentum shifts before the AMM repricing catches up.
Continuous YES + NO < 1.00 scanner. When the sum drops below par after fees, the bot fires simultaneous buys both sides and merges into a risk-free position.
Spread arbitrage on equivalent contracts between Polymarket and Kalshi. Opposing positions taken when the spread clears combined venue fees.
Contrarian value play against late-cycle Kalshi mispricings. Realised-vol gate plus RSI confirmation; small clip-size, frequent entries.
Quantitative models in play
- Sigmoid fair-probability with realised-vol shaping
- Kelly criterion, Monte-Carlo calibrated, fractional
- Particle filter for hidden-state regime tracking
- Importance sampling tuned to SNIPE tail mass
- Jump-diffusion Merton model for tail entries
- Kyle’s lambda on order-book quality
- Brier score continuous self-calibration
Operational footprint
1,000+
Live-tested trades (Kalshi V6.2)
4 assets
BTC / ETH / SOL / XRP
7
Strategies live
Telegram
Remote kill + status
Stack: Python 3.13 · asyncio · Docker on AWS EC2 · Telegram for remote kill and status · Chart.js dashboard with Cloudflare Tunnel exposure. Kalshi via RSA-signed REST/WS keys; Polymarket via CLOB + Polygon RPC.