Stephen Robinson
Thirty years architecting data at banking scale — now building the production AI that runs on it. I bring 30 years of banking-grade data architecture to production AI: I designed and shipped the MOTLogs estate — five live production systems spanning a voice booking agent, a self-grading agentic RAG knowledge base, a predictive MOT-log analytics engine, a 136M-record risk-scoring engine, and a code-intelligence platform exposed to AI agents over MCP.
Track record — Nomura, 2009 – present
Global Head of Risk IT Reporting. Before MOTLogs, three decades making data dependable at the scale and regulatory rigour a global bank demands. This is the foundation the AI work is built on — not a pivot away from it.
The estate — five production systems, architected end to end
Customers ring a normal phone number and talk to an AI receptionist that books appointments against a live calendar, answers FAQs from the garage's knowledge base, takes messages, and transfers to a human when it matters. Every conversation flow is data-driven, so a new garage is configured from an admin UI — no code change.
A tenant-isolated knowledge base that answers customer and staff questions with retrieval-augmented generation and grounded citations. When confidence is low it escalates to a human over Twilio Video, and it continuously learns from ratings, detects drift, and flags weak answers for a governance reviewer.
MOTLogs ingests vehicle MOT test logs from the DVSA history API and user uploads, turns each test's failures and advisories into a semantic narrative, and predicts the likely outcome of a vehicle's next test by comparing it against cohorts of look-alike vehicles. It powers a garage-facing web app and an analytics API for compliance and test-quality review.
Scout ingests the entire UK DVSA MOT dataset — ~49GB, 136 million records — and gives every vehicle a deterministic 0–100 risk score from an eight-stage pipeline that weighs defect patterns, mileage, ownership behaviour and model-specific failure rates. A registration lookup then returns an instant, mechanic-grade risk read for auction buyers.
OpenBrain parses every repository in the estate into a hybrid semantic + structural index — a code knowledge graph — and serves it to AI agents over MCP. An agent can ask where a function lives, what calls it across repos, why it exists, and what a change would break — and get grounded, cited answers instead of guesses.
Capabilities mapped to 2026 demand
A Wiki knowledge base with hybrid dense + full-text retrieval, a cross-encoder reranker (bge-reranker-base, int8 ONNX), GraphRAG, and grounded citations carrying source URIs into every answer.
A production MCP server (mot-ledger) exposing 25+ tools — semantic search, dependency graph walks, 'why does this exist' provenance, invariant guards, refactor planning — that turn a whole codebase into a queryable substrate for AI agents.
Drift detection, self-evaluation runs, RLHF reweighting, active-learning queues and golden-set ranking validation — wired as hosted services with a governance portal a human reviewer signs off in.
A Twilio ↔ ElevenLabs ConvAI orchestrator with a custom-LLM contract, spike-driven engineering, and explicit, measured SLA budgets (p95 latency gates) — built on telephony I'd already shipped at PhoneBounce.
The instinct from 30 years of regulated banking data: multi-tenant row-level security, DSAR/privacy, a guardrail hook that blocks agents answering before they consult the source of truth, and audited overrides.
Reproducible Hetzner infrastructure (idempotent bootstrap, Tailscale, disaster-recovery runbook), Stripe billing, nginx/Kestrel, and commit-driven re-indexing — the unglamorous parts that make AI a business.
How I work
Thirty years delivering regulated reporting at a global bank means I design for governance, lineage and recovery from day one — the questions a board will ask are answered before they're asked.
Embedder choices A/B-tested, latencies measured live, SLA budgets re-based against reality with a written justification. Decisions leave a trail you can audit.
I orchestrate AI agents to do real engineering — a workflow-driven QA loop ran parallel testers, adversarially verified findings, auto-fixed and redeployed, driving confirmed bugs 13 → 2 to convergence.
Lose the server, run one idempotent bootstrap, you're back. Infrastructure as a runbook, not as tribal knowledge in one person's head.