Work

Systems, experiments, and evidence

Selected public work with inspectable sources, bounded claims, and explicit limitations. Each case shows the problem, the system boundary, what I built, how it was evaluated, and where the evidence stops.

• Engineering the Agentic Stack

  Problem: Agent reliability is often treated as a model-choice problem even when failures cross reasoning, memory, tools, permissions, runtime, and release control.

  Contribution: I developed a six-part production architecture and the Market Analyst Agent reference implementation that connects those layers.

  Evidence: Read the series · Inspect the source

  Decision: Treat the harness and its independent gates as part of the product, then make each added control earn its cost through ablation.

  Limitation: This is a reference system and engineering method, not a published production benchmark.

• Traces to test suites

  Problem: Answer-only grading misses skipped tools, unsafe arguments, loops, and premature success.

  Contribution: I wrote a trace-to-regression method for production agents and built a runnable RAG evaluation harness spanning retrieval, filters, reranking, generation, judges, and latency.

  Evidence: Read the agent evaluation method · Run the RAG evaluator

  Decision: Version failure traces as test cases and combine deterministic trajectory checks with calibrated judges only where interpretation is required.

  Limitation: The public evaluator uses SciFact and bounded examples; its results do not transfer automatically to another corpus or model.

• TypeScript Agent Service

  Problem: Python ML engineers need to ship and review production-shaped TypeScript agent services without relearning backend engineering from zero.

  Contribution: I built a pnpm monorepo with a streaming Hono API, validated tool loop, Drizzle storage, worker, MCP server, observability, and 40 tests.

  Evidence: Read the implementation guide · Inspect the source

  Decision: Map familiar service boundaries—Pydantic, FastAPI, queues, storage—to their TypeScript equivalents while keeping schemas strict at every tool boundary.

  Limitation: It is a local reference service, not a hosted product or a claim about a specific employer system.

• Search ranking stack

  Problem: Retrieval stacks hide which stage improves relevance and which stage merely adds cost.

  Contribution: I built a five-stage demo covering BM25, dense retrieval, reciprocal-rank fusion, cross-encoder reranking, and LLM listwise ranking on sampled Amazon ESCI data.

  Evidence: Read the staged analysis · Run the source

  Decision: Measure every stage independently and retain it only when the relevance gain justifies latency and operating cost.

  Limitation: The corpus is a sample and the article marks the stale LLM row as non-evidence.

Inspect the evidence directly: run a lab or browse the Blog by agent systems, retrieval and language, or model engineering.