AI Engineer building production LLM and RAG systems — and the evaluation infrastructure that keeps them reliable. Based in the UK.

I run hgray.agency, a product studio where I ship my own products and client work across interview tech, healthtech, fintech, restaurant ops, femtech, edtech, logistics, and developer tooling. Before AI engineering I spent seven years in quality engineering across finance and healthcare — which is why everything I ship arrives with tests, evals, and observability attached.

AWS Certified Cloud Practitioner · Verified n8n Creator · PRINCE2 Foundation

A desktop app for serious interview preparation: a structured prep workspace, AI mock interviews with voice, and a knowledge base that keeps every generated answer grounded in your real experience. Local-first — your data lives in SQLite on your machine, you bring your own AI key, and prep works offline.

• Live: ofarwise.com · public release feed: ofarwise-releases
• Electron + React 19 + TypeScript (strict), better-sqlite3, Anthropic Claude with prompt caching, Deepgram speech-to-text
• 2,000+ tests across Vitest and Playwright, shipped with auto-update. Source is private.

Production-grade RAG pipelines with evaluation baked in. A framework-agnostic CLI that scaffolds production RAG systems, runs continuous evaluation as a CI/CD gate, and scores any pipeline against the RAG Maturity Model (RMM-0 to RMM-5). Golden-set regression evals, drift detection, guardrails and PII scanning, per-stage OpenTelemetry spans, and an MCP server.

Six packages on npm and PyPI — 7,500+ combined downloads. TypeScript CLI plus a Python pipeline, delegated through a subprocess bridge. MIT licensed.

• Docs: rag-forge-docs.vercel.app
• Website: rag-forge-site.vercel.app
• Install: npm i -g @rag-forge/cli

B2B clinical-intelligence platform for African primary care. One engine — RAG over curated, guideline-grounded medical knowledge with four-level triage, a deterministic red-flag safety layer, and multi-LLM validation judges — powering a multi-tenant hospital platform and a FHIR R4 API. Built to the Nigeria Data Protection Act 2023. Platform live at pearmedica.com; source is private.

A pre-hyperopt screening framework for testing whether a candidate trading signal carries information over a baseline — before you spend cloud hyperopt budget finding out it doesn't. Four mechanical screens (coverage, lift, Jaccard, information-split) run in seconds on a laptop and produce a JSON verdict. Methodology applies to any signal hypothesis on any asset class; the repo ships a worked crypto case study where four well-known strategies failed pre-committed criteria — one of them killed by SIA in ~3 seconds, before any hyperopt ran. Apache 2.0.

• Blog write-up: Four crypto strategy failures, and the tool that caught one in 3 seconds
• Install: pip install -e git+https://github.com/hallengray/signal-independence-audit

A follow-up case study within SIA: a regime-switching strategy that combined two strategies which had already failed on their own, riding a trend-following signal in bull markets and buying dips in sideways ones. It became the first signal in the whole project to clear the SIA screen, then lost money out-of-sample anyway (best run about -9% over 16 months, profit factor 0.78, Sharpe roughly -0.23). The regime switching worked correctly. One of its two legs was simply unprofitable on unseen data, and the other could not carry it. A clean reminder that clearing the screen means "worth the expensive test," not "this works."

• Full kill record: case-study/11-mss-kill-adr-0027.md

Commercial and client products built end-to-end under Hgray. Code is private; some are live, others are in development for clients.

NosHQ (live) — Multi-tenant restaurant POS: real-time order management, kitchen display system, floor planning, inventory, and full UK compliance (VAT, allergens, tip law). In production across two venues. Built with Next.js 16, Supabase, and shadcn/ui.

Correx (live) — Find the cheapest way to send money abroad. Compares live quotes from 20+ providers across UK to Nigeria, UK to Ghana, UK to India, and more.

MumWell — Molly's Wellness Club (live) — A wellness platform for postpartum women across the UK. Personalized support, nutrition tracking, community engagement, and expert guidance.

AkwaabaTech (client, in development) — Ghana's first comprehensive educational platform combining affordable vocational training (barbing, carpentry, fish farming) with digital skills courses, paid via mobile money.

VanGo (client, in development) — Uber-style van hailing for bookings, moves, and deliveries. Instant booking, real-time tracking. Built on React 19, Next.js, TypeScript, and Tailwind CSS 4.

AuraTest (in development) — AI-driven test management platform with autonomous test orchestration. An intelligent Test Manager handles project setup, test planning, and reporting while testers focus on manual execution.

• n8n-nodes-twelve-data — n8n community node for the Twelve Data financial-markets API: 161 operations spanning market data, 91 technical indicators, and fundamentals. Reviewed and verified by the n8n team; 1,800+ downloads on npm. (source)
• supabase-auth-starter — Reusable Next.js + Supabase authentication template with OAuth.
• football-edge — Big-5 football value-betting research tool: per-league trained models, vig-stripped fair prices, fractional-Kelly staking, AI-explained picks, and an honest backtest record. Live demo (password-gated).
• EaseMyBP — Blood pressure tracker with daily readings, charts, diet and exercise logging. Live.

AI: Anthropic Claude, OpenAI, Google Gemini, RAGAS, DeepEval, LLM-as-judge evals, OpenTelemetry + Langfuse, MCP, pgvector Frontend: Next.js 16, React 19, TypeScript (strict), Tailwind CSS, shadcn/ui Desktop: Electron Backend: Supabase (PostgreSQL, Auth, Storage, Edge Functions, Realtime), Node.js, Python (FastAPI) Infra: Vercel, AWS, Docker, GitHub Actions, Turborepo, pnpm workspaces, uv

• Website: hgray.agency
• LinkedIn: linkedin.com/in/femi-adedayo-h44
• X / Twitter: @hallengray
• GitHub: @hallengray