Runnable references that show how I approach retrieval, evaluation, failure handling, APIs, product interfaces, and delivery. Synthetic fixtures are labeled and never presented as real-world usage.
A TypeScript support operations console with cited streaming responses, Zod-validated structured actions, human approval gates, retry and recovery behavior, and a transparent synthetic reliability dashboard.
One refund policy implemented twice, as a tool calling loop written by hand and as a LangGraph state graph, then benchmarked against a deterministic model across three scenarios. Identical behaviour, 97 lines against 50, with the benchmark enforced as a CI gate.
Offline evaluation for citation validity, expected-fact coverage, abstention behavior, latency SLOs, and estimated model cost. The same JSONL dataset produces the same result locally and in CI.
A runnable retrieval reference covering document ingestion, chunking, deterministic embeddings, lexical and vector ranking, reciprocal rank fusion, cited answers, no-context abstention, offline evaluation, Docker, and CI.
A polished social media post generator for LinkedIn and X with dark/light mode, multi-language support, instant image export, and a clean content authoring workflow.
The static platform behind this site: responsive navigation, theme handling, accessible contact flows, serverless integrations, canonical routing, and automated browser checks.
Open Relay Desk, generate a cited reply, inspect proposed actions, approve a safe action, then review the synthetic reliability dashboard.
Open live demoReview the golden-set contract, failure taxonomy, non-zero CI gate, and deterministic HTML report in LLM Reliability Lab.
Open repositoryFollow ingestion through lexical and vector ranking, RRF, citations, abstention, and offline Hit@3/MRR fixtures.
Open repository