VIGIL — Multi-LLM council architecture
VIGIL is an AI meeting copilot. Its core design principle: no single model is trusted alone. Every intervention is drafted, reviewed and voted on by a council of models from different families before it reaches the user.
The council pipeline
- Ingest — a real-time meeting transcript flows into the AI Worker Collective.
- Primary worker — Claude Sonnet, specialized per role (CFO / Tech / Legal / Product advisor), drafts a proposed intervention.
- Reviewers — GPT-4o and Gemini Flash score the draft on six dimensions in structured JSON. Different model families are used deliberately, to avoid bias collapse.
- Consensus vote — a weighted vote must pass a
66%consensus threshold and an80%readiness gate. - Chairman fallback — if consensus fails, a chairman model synthesizes the final answer.
In parallel, the Behavioral Pattern Engine (550+ patterns across 50 ensembles) reads each speaker's transcript window and emits live signals: stress, evasion, learning agility, persona alignment.
Stack & validated numbers
| Layer | Technology |
|---|---|
| Runtime / API | Node.js 20+, Express |
| Memory / RAG | Supabase Postgres + pgvector |
| Cache / queue | Upstash Redis |
| Council models | Claude Sonnet + GPT-4o + Gemini Flash (direct APIs) |
| Live audio | Recall.ai + LiveKit, Groq Whisper STT, ElevenLabs TTS |
| Billing / auth | Stripe, Supabase Auth |
Validated proof-of-concept results: $0.008 per council run (37× under the $0.30 viability threshold), P95 latency 10.0 s (under the 15 s live-whisper budget), 10/10 runs flagged the correct issue in the test scenario, with real reviewer diversity (70-point harm-risk variance between reviewer models).
VTLVS — Universal wireless charging pad
VTLVS (pronounced Atlas) is a single hardware SKU that wirelessly charges every class of electric vehicle. The design bet mirrors USB-C: win by being one abstraction over every device.
| Specification | Value |
|---|---|
| Form factor | 700 × 700 × 200 mm |
| Environmental rating | IP68 waterproof · IK10 impact |
| Power range | 250 W (e-bike) → 22 kW (SUV) — 90× dynamic range |
| TX coils | 4 nested coils, r = 360 / 280 / 200 / 80 mm + solid-state relay |
| Vehicle classification | Weight-first, < 200 ms — no tag, no app, no connectivity |
Four revenue models, one pad
- B2C retail — direct sale for homes and buildings.
- B2B parking — Hardware-as-a-Service for parking operators.
- B2G municipal — curbside charging concessions for cities.
- B2B2C hospitality — charging as an amenity, billed via SaaS.
One SKU across all four collapses stock, supply chain and support — the strategic insight behind the product.
JobBoat — Career platform
JobBoat merges three experiences that today live in separate apps: inspiring career content, professional networking, and job discovery with genuine profile understanding.
- Web platform plus native mobile apps (Expo + React Native, iOS & Android).
- Profile-aware recommendations — matching driven by deep profile understanding rather than keyword search.
- Behavioral engine lineage — the pattern-analysis technology developed for JobBoat became the foundation of VIGIL's Behavioral Pattern Engine.
- GDPR-first — user data powers recommendations and is never resold.
Our delivery method (services)
Every client engagement — automation, agentic installation, systems, web or data — follows the same auditable path:
- Diagnose — free consultation; we map workflows, data and constraints, and rank automation candidates by leverage.
- Design — a written architecture and rollout plan: scope, guardrails, costs, success metrics. You approve before any build.
- Deploy — short iterations, human-in-the-loop gates on anything customer-facing, audit logs on every automated action.
- Operate or hand over — monthly operation by us, or a documented handover your team can run without us.
Design principles we do not compromise on: deterministic routing (rules decide which system handles what — not vibes), prepare-only vs. execute separation (analysis environments can never send or write to production), and cost governance (per-run budgets, monitored).
FAQ
Can we self-host what you build?
Yes. Client systems are built on your infrastructure or your cloud accounts, with documentation and handover included by default.
Which AI models do you use?
Whatever fits the task and budget — typically Anthropic Claude, OpenAI GPT and Google Gemini via direct APIs. For review-critical steps we deliberately mix model families.
Do you work in English or French?
Both, natively. Contracts, documentation and support available in either language.
How do we start?
Book a free consultation — we reply within 24 hours.