Hand me the system nobody wants to own.

I designed and delivered the orchestration architecture: an orchestrator routing requests across specialised agents, provider failover so one vendor outage cannot take the feature down, and a MongoDB layer carrying document retrieval, agent execution logs and a response cache that cuts the cost of repeat inference. The provider was not chosen by reputation — five models were benchmarked on cost, latency and output quality, and the numbers picked one. In production it serves 100K+ users, with latency down 90% and cost down 40%.

Stack: LLM orchestrationMulti-agent routingProvider failoverMongoDBNestJS

A core query took two seconds and nobody could say why. EXPLAIN ANALYZE said it plainly: the divergence between estimated and actual rows exposed a sequential scan chosen over an index that already existed. I refreshed the statistics, restored correct index usage, tuned the autovacuum thresholds so it would not come back, and wrote the method down for the team — the fix is worth less than the second time somebody finds it themselves.

Stack: PostgreSQLEXPLAIN ANALYZEautovacuum

Revenue, cost and cash-flow reporting used from the operations floor up to the owners, handed over with nothing. I reverse-engineered it, assessed the debt and set the migration strategy alone, then stabilised it on three fronts: killed the crashes, traced and corrected the figures it had been reporting wrong, and cut load times. The migration ran incrementally while the system stayed in daily use — no feature freeze, no big-bang rewrite, no interruption to financial reporting. Promoted within six months.

Stack: ReactTypeScriptKubernetes

A frontend Docker image nobody wanted to touch, with no prior documentation to work from. I diagnosed the cause rather than trimming layers at random, then rebuilt the pipeline around multi-stage builds. Smaller artifacts dropped registry and transfer costs by over 60% on top of the build time.

Stack: DockerMulti-stage builds

Thirty questions over fifty Python Enhancement Proposals, with the ground truth recorded as character spans rather than chunk ids, so changing the chunking cannot quietly rewrite the answer key. Dense, BM25, hybrid and a cross-encoder reranker are compared question by question: the reranker lifted recall to 0.833. Every answer quotes its sources, and code resolves each quote to characters in the document — 51 of 52 were found verbatim. CI replays BM25, dense, hybrid and every recorded answer with no key and no network call, and fails on any change.

Stack: PythonPostgreSQLBM25Gemini embeddingsCross-encoder rerankingAstro

citation viewer(opens in a new tab)github(opens in a new tab)

An approval gateway for AI agents. An MCP proxy intercepts tools/call and holds the risky ones until a human signs, with a declarative policy engine so approval is by exception rather than by default, and elicitation and sampling relayed back in both directions. Every decision lands in a hash-chained audit trail that verifies offline. Agent code does not change to adopt it — the proxy is the integration.

Stack: MCPHonoDrizzlePostgreSQLApache-2.0

useseal.dev(opens in a new tab)github(opens in a new tab)npm(opens in a new tab)

An embeddable Pix checkout that runs on the merchant's own domain, inside a Shadow DOM, with no runtime dependency in the browser and no request beyond the merchant's own backend. No credential can reach the bundle, because the SDK takes functions and not keys. There is no onSuccess, because the truth of a payment is a signed webhook and never the browser. Illegal states are unrepresentable and the promised amount is an invariant.

Stack: TypeScriptShadow DOMPixZero runtime deps

A platform that scores client health from Google Ads and Meta Ads signals, learns the weights per client rather than applying one model to everybody, and writes a weekly brief an account manager can act on. Multi-tenant with row-level security, Stripe billing gates, and observability built in from the first deploy rather than added after the first incident.

Stack: Next.jsMulti-tenant RLSStripeGoogle AdsMeta Ads

A history of technology told through eight objects, from a struck stone to machines that are learning to judge. One WebGL point cloud is reassembled into every era's form, the interval between two eras sets the height of the silence before it, and a single accent walks the spectrum as you read. The heavy layers wait for the reader's first move, which is why the artefact survives the score.

Stack: three.jsGLSLGSAPLenis