I build LLM applications around business context, tools and human review.

Focus: RAG and retrieval, structured outputs, MCP and scoped tools, evaluation and caching.

Stack: Python · TypeScript · Node.js · FastAPI · PostgreSQL · SQLite · Docker · GitHub Actions · Git · Linux.

Evidence tooling: GitHub Profile Audit — read-only checks for public repository and profile-README consistency, with tests and CI.

Selected open upstream PRs (checked 2026-09-30 UTC; none are merged claims): LlamaIndex #23299 validates non-positive workflow iteration budgets; NanoCoder #1528 makes an LLM-summary fallback visible; Pydantic AI #8955 clarifies agent persistence and sandbox docs; Clinical Deep Research #91 replaces publishable-harness print() output with structured logs; Haystack #12986 fixes literal tiktoken markers crashing TiktokenCounter; Jev RAG #9 adds Windows CLI smoke coverage (maintainer changes requested; still open). These are selected open submissions, not merged contributions.

Merged upstream: mcp-memory-service #1334, #1343 and #1344, all merged into main with maintainer merge commits; Clinical Deep Research #90 adds offline regression coverage for long ClinicalTrials.gov query sanitization, and #92 aligns critique output with the current schema (both merged 2026-09-30 UTC).

The #92 validation also surfaced an invalid mypy override; the separate CDR #95 fix corrected it and credits the report from #92.

Technical discussions: Haystack #12969 proposes a fail-closed context-budget guard for RAG; Jev RAG #5 proposes an opt-in OCR adapter with page-aware citations and explicit privacy boundaries; mcp-memory-service #1345 proposes idempotent event-log sync and embedding-consistency invariants. RFC author filhocf confirmed the invariants were incorporated into the v0.3 draft (design-only); the linked discussions and issue are proposals, not merged implementations.

GitHub snapshot 2026-09-30 08:49 UTC: 30 public repos including this profile; 249 commits, 46 pull requests, 2 reviews, 2 issues and 355 total contributions in the 365 days ending at the snapshot time, verified against GitHub's public profile and contribution collection. These are dated counters, not live values. The separate activity/language panel is a 2026-09-28 snapshot; language shares use bytes from public owned repos, excluding forks and this profile, and are not proficiency scores.

Turns shipment exceptions into sourced next actions an operator can review.

Python · FastAPI · SQLite · MCP. Simulated operations · scoped access · human approval.

The separate Agent Gym prototype checks evidence-grounded tool-use proposals against synthetic cases, including policy citations, tenant scope and forbidden-action attempts. It is offline and does not claim live-model quality. Implementation and limits.

Tests whether a cached answer is still safe after its context or permissions change.

Python · Evaluation · Cache invalidation. Reproducible cases · exact / semantic / dependency-aware.

Brings source retrieval, structured drafting and Word/PDF export into one review flow.

TypeScript · Node.js · Zod · RAG. In development · legal-source validation remains ongoing.

Email · Repositories