Agents That Survive Contact With Production.

I broke a local café's AI reservation agent and talked it into refunding a booking it never should have. I built this harness as the proof-of-concept and gave it to the software agency that had shipped that agent — they had no dedicated AI engineering team. They have used it as the template for their new projects since.

A nine-layer refund agent built specifically to be attacked. LangGraph, Qdrant, custom MCP tools.

88.6% pass across 229 adversarial cases · 98.7% injection resistance · 100% jailbreak resistance

Built for a profitable TMS company to cut cost and modernise how the organisation actually moves documents around. It was saving them time and adding revenue inside the first week.

Hybrid retrieval over logistics and TMS documents — dense plus BM25 with reciprocal rank fusion, four deterministic guardrail layers, and an LLM-as-judge evaluation framework, traced in Langfuse.

91.9% answer accuracy · $0.003 per query

Designed as a proof-of-concept for a medtech startup. The PoC was accepted and taken on by their own team for further development and use.

Procurement intelligence briefs for medical substances — product profile, clinical use, market, manufacturers, regulatory status, sourcing and risks, with every claim backed by a verified source URL. Four-step safety escalation, versioned prompts, per-call cost accounting, observable traces.

Built to visualise the jungle in my head — the closed-eye-dream version of it. It comes out of a love of deep forests and the novels of Jim Corbett and Kenneth Anderson.

Turns hunting and natural-history novels into illustrated, evaluated bestiaries. Five-stage async pipeline — ingest, extract, validate, enrich, render — with LiteLLM provider abstraction (Gemini primary, Groq fallback, Ollama offline), Wikidata validation, and AI-illustrated PDF export.

A wearable-first app for smartwatches and bands, built on minimal interaction, haptics, privacy, and a close connection between two people.

End-to-end encrypted messaging. X25519 pairing ceremony, HKDF-derived subkeys, AES-256-GCM. Security review found the six-emoji verification phrase grindable at 2^36 by the relay server — precisely the adversary the design distrusts. Mitigated with iterated hashing.