AI Engineer — verifiable LLM systems

AgentA — self-improving engineering harness

An autonomous Claude Code system — specialised subagents, skills and hooks — that improves its own code behind a verification gauntlet. Four gates always run: parse, unit tests, a benchmark delta, and a fresh-context critic that judges the diff with no memory of how it was produced. Three more (property tests, mutation testing, a sealed holdout) arm only when a task configures them.

A skipped gate is logged as skipped and never counted as a pass

Inner research loop ported from Udit Goenka’s autoresearch (MIT, after Karpathy); the meta-improvement architecture and verification stack are my own.

Financial-Statement Extractor

Turns a government audit statement (PDF) into clean structured JSON, then re-derives every total to prove the extraction is correct, with a golden-file regression test. The model never does the arithmetic — tested code does.

25 checks, 0 exceptions — inject one wrong figure and it’s caught

GrantLedger

Full-stack B2B SaaS that auto-categorizes nonprofit grant spending into 2 CFR 200 budget categories, tracks budget-to-actual per grant, and generates audit-ready compliance reports, with QuickBooks/Xero integration.

Built and running locally, not yet deployed — source is public

Legal RAG with page-level citations

A retrieval system that lets attorneys query a legal-text repository and get answers carrying precise page-level citations, so every claim can be checked against the page it came from. Retrieval plus LLM synthesis, built and deployed at a law firm and used in live case work.

An answer without a citation to a real page is not an answer

Built for a client, so the corpus and the code stay private. Happy to walk through the retrieval and citation design in a conversation.

NeuroFour — Connect 4 strength-per-byte

Connect 4 is a solved game, so perfect play is a known, fixed target. NeuroFour ranks 20 Connect-4 agents by strength per byte, scored against an exact solver. Nineteen fit the hard 5M-FLOP-per-move budget and are eligible for the headline; the twentieth is that solver itself, which needs 50M FLOPs and is listed but over budget.

Zero leads optimality, ladder Elo, and NeuroFour Score among the 19 under-budget agents, at 0 bytes, ahead of all 14 learned-net entries (6 distinct weight files, largest ~24 KB): ladder Elo 754 vs 631 and NeuroFour Score 96.45 vs 74.25 — a score that divides strength by a size penalty — optimality 0.960 vs 0.957 by a single move in 300. Each “vs” is the strongest learned entry on that axis (optimality is a tie); the four entries behind them load just two weight files. The cost: Zero spends 4,999,028 of the board’s 5M-FLOP budget — 99.98% of it, 18th of 19 on FLOPs/move (and 4th on soundness, 15th on latency).

Free-tier API — first move may take ~30–60s to wake.