I build the verification layer for AI systems — the gates, tests and CI that decide whether a system is actually right, and that fail loudly when it isn't.

Open to AI engineering roles where the evaluation harness is the job. Based in Riemst, Belgium · open to remote roles · employment or contract.

📧 bakerstreetbandit@zohomail.eu

Every number below came out of a command you can run. The command sits next to the number.

• elohim — next measurement track (mutation survival, shard pinning), landing daily
• harness — v1.1.0 wheel and CI readiness floor; package release pending
• terminal221b — release tagging and install hardening

Six instruments measure hard mathematics, pin every result they claim, and refuse to pass if anything moved — including the instrument itself. Break a file on purpose: the gate names it.

python3 tests/test_all.py   # → ALL_SKILLS_PASS; verdict FAIL + PIN DRIFT on any tampered copy

81 facts · 107 checksum-pinned values · 0 unclassified · 6/6 seeded tampering caught · 43 commits, 22,196 non-blank lines, measured at a047ccd

scan → task → verify → pr. Protected-path writes and secret reads are blocked before an agent's edit lands; a PR refuses to open unless the verdict is PASS — a skipped review counts as not passed. The evaluation runs the same agent raw and harnessed against a hidden acceptance suite: 11/13 graded PASS, and the unflattering number is published too (scope violations: raw 3, harnessed 4).

pip install git+https://github.com/BoozeLee/harness.git   # 1.1.0
pytest                                                     # → 88 passed

88 tests · 11 subcommands · 48 commits · scan --fail-under 70 exits 1 below the floor · wheel + sdist build via uv build

TypeScript CLI with a Rust ratatui TUI and an Expo client. Context is bounded by path, not by a token budget; untrusted output is redacted before it reaches a stored transcript; the test modules are the containment properties (path-guard, boundary-drift, scope).

npm ci && npm test    # → 385 passed
cargo test            # → 60 passed

445 tests total (385 TypeScript + 60 Rust) · 55 commits · 16,210 non-blank lines

A GitHub Action that diffs an MCP server's tool contract across releases, so a renamed or narrowed tool fails CI instead of failing in production. Untrusted tool text is escaped before it reaches a PR comment.

npm ci && npm test    # → 29 passed

29 tests · 1,247 non-blank TypeScript lines

Also public: repotruth — CI that reads the repository, not just the diff: signed-webhook verification (replay → idempotent 200, unsigned → 400 and nothing queued, tampered body rejected), one audit core exposed as CLI, MCP server and fleet dashboard. 147 tests across 32 suites.

Private, available to show on request: capo (349 tests across five packages) · superbrain (298 passing tests) · beehive-studio (54,550 non-blank lines).

Give an agent a task and it will report success whether or not it succeeded. The interesting work is the part where you check. So I built the checks I wanted to exist: instruments that re-derive their own numbers, gates tested by being broken on purpose, a harness measured against the thing it replaces — including where the measurement is unflattering. That is the work I want to be paid for.

No degree or certifications, and no prior employment — this would be my first role. Bakery-street-project is my personal project, not a company. Every public repo has 0 stars and no users; there is no adoption story. Docker, Kubernetes and Go appear nowhere in my code. Everything above is verifiable by cloning the repository and running the command shown.