I design, ship, and validate real systems: agentic pipelines, payment integrity, AI authoring tools. Then I make them usable, for the people and the models downstream.
Nine Claude Code loops write, review and ship the code unattended; two structural GitHub gates decide what reaches production, never the agents' good behaviour.
Production AI systems, hardened software, and the systems thinking that connects them.
The first live trial produced zero data, and the reported symptom — "the link expired" — was false, sent by our own error page. Reconstructing the morning from production logs found a URL wrapping mid-string on a phone and a text detector that stops at the line break. Every fix that kept a link kept the failure, so I shipped an entry path with no link in it.
A weekly agent walks the product at two viewports and judges it from what it captured — screenshots, page text, console errors, every confirm dialog. The harness asserts nothing, because one that fails on an assertion only teaches you what its author already knew. Every design decision in it targets the same failure: the silent pass.
A production pilot running inside a real estate agent's week: buyers answer 7 questions on their phone in under 90 seconds, the agent gets a ranked Monday call list. The engineering that mattered was refusing the features that would have made the data untrustworthy, and pre-registering a counterfactual so the pilot was allowed to come back with an unwelcome answer.
Reviewing a 2,000-line agent diff is theatre. I moved the review upstream: requirements in WHEN/THEN form are canonical, agents work as deltas against them, and previews are structurally unable to touch production. The rule file records its own near-misses — including two archive bugs that would have silently deleted live requirements.
Firestore rules deny every client read and write, so authorisation lives entirely in server code with no database backstop. I built the three-layer auth chain and a suite that forges nothing: it mints real session cookies through the app's own sign-in route, then proves all 16 mutating actions are denied across the boundary and write zero bytes.
An AI writing pipeline where one agent writes and a separate one grades it — because nothing should grade its own work. A three-part pass/fail gate, enforced by a dependency-free validator and calibrated against human-graded examples, keeps polished-but-wrong content from getting through.
A founder's billing prototype had to start taking real money — and a stale test-mode ID had already silently charged a card $138 with no record of it anywhere. I locked the billing data down at the database, made the webhooks heal their own bad data, and wrote 142 tests from zero.
Label a GitHub issue, and a locked-down, automated Claude Code run opens a draft pull request and a preview. The workflow handles git, not the AI. Trust-gated and hardened against prompt injection. I debugged it across a live run of 4 pull requests, then honestly parked it behind a product pivot.
A personal wiki the LLM writes once and keeps current, instead of rebuilding an answer every time you ask. It summarises, cross-links, and flags contradictions across 100 linked pages (none written by hand) drawn from 66 sources. A deliberate alternative to query-time RAG.
Worked out how a user could cheat a SaaS billing system from the browser — give themselves a paid tier, inflate their own credits — and shut it down at the database. Every protected field is locked, and each edit is checked against what actually changed. Proven with 59 allow/deny tests against the real rules in the Firebase emulator.
AbletonMCP, an open-source MCP bridge, could write MIDI clips but never read or truly modify one — so an AI agent had to regenerate each clip from memory, blind to manual edits. I designed and shipped two MCP tools that close the read → modify → write loop, with a 13-test suite. Submitted upstream as PR #106 (+612 / −0).
A three-step agent pipeline that reads an API release, works out which docs it affects, and drafts the updates. A human reviews them before they reach developers.
An LLM that scores submissions against a clear rubric, with checks that catch judgments it can't back up. It runs at a fraction of the cost and time of grading by hand.
A diagnose-and-fix campaign on a ~110K-record address dataset. A 5.8% per-run success rate looked like a broken geocoder — but the real problem was the shape of the addresses. I wrote the normalizer that recovered 47% of the failing backlog and lifted whole-dataset coverage to 98.3%.
A 5-person program that ended the duplicate docs hurting search and AI retrieval. One source of truth per concept, with a reusable 0-4 classifier deciding what stays. ~2,660 lines consolidated across 8 domains, with one site retired.
A five-stage serverless pipeline behind the in-app report-a-bug button: a locked-down screenshot upload, a private GitHub issue, a signature-verified webhook when the bug is fixed, and a two-pass cleanup of old data. Shipped to all users, with about 1.3 lines of test for every line of code.
The CI orchestration that keeps a fragile ~40-job scraping fleet reliable. One failing job can't sink the rest, large-file commits stay intact even when the main branch moves mid-run, and failures are sorted into infra-vs-bug so only the right ones auto-retry.
A conversational AI that mentors through Socratic, in-character dialogue. An experiment in how far one well-designed prompt system can carry the whole experience.
The stack I design, ship, and validate with.
I don't just call models. I build with the full agentic toolchain. What sets the work apart is that I actually run the stack, not just the API.
I design, build, and ship production AI systems: agentic CI pipelines, billing systems that don't lose money, AI authoring tools. Then I make them clear to the people and models that depend on them. Building something well and explaining it clearly don't always come from the same person, and I work at both.
As a technical co-founder at storipro, I shipped the production billing layer, hardened the data-security model, and built an autonomous issue-to-PR pipeline. At commercetools, I architected an orchestrator-agent authoring system with an eval harness and drove a single-source-of-truth documentation program across the platform. On my own time I keep building. An LLM-maintained knowledge base, an MCP control layer for Ableton Live.
The edge behind that is a decade in education before this: the craft there is earning enough trust that people actually engage. I build the same way, for the humans and the models that have to use the output.
If you're building applied-AI systems, or you need someone who can architect them, ship them, and make them land, I'd love to connect.