Projects

Applied AI systems, built in public.

Production-grade extraction, retrieval, and automation systems. Every project ships with the code, architecture decisions, benchmarks, and failure reports.

Quantara — product extraction backend

Live

Self-hosted backend for e-commerce research. Paste a product URL or upload a PDF catalog; Claude returns a typed JSON product (title, SKU, brand, price, specs, variants, images, availability) and joins a batch that exports to a multi-sheet Excel workbook with price-history deltas and margin calculations. No per-vendor selectors — text path with vision fallback when extraction is thin or blocked. Stealth Puppeteer, typed error codes, SQLite price history.

  • Claude Sonnet 4.6
  • Tool use
  • Puppeteer
  • Node.js
  • SQLite
  • Vision fallback

Regulatory intelligence RAG platform

Currently shipping

Retrieval pipeline over public regulatory documents: ingestion, semantic chunking, embeddings, pgvector indexing, and a Next.js research interface with source-grounded citations. Includes golden datasets, regression testing, and retrieval benchmarks to prevent prompt drift.

  • RAG
  • Embeddings
  • pgvector
  • Next.js
  • Evals

Personal knowledge MCP server

Shipping next

Model Context Protocol server exposing structured personal data to local AI clients through resources, tools, and prompts. Built with stdio transport, protocol tracing, and automated validation using MCP Inspector.

  • MCP
  • TypeScript

Workflow automation agent

On deck

A production workflow agent built directly on Anthropic tool use. Every decision, tool call, retry, and failure is logged, evaluated, and documented. Target workflow: document processing and action generation.

  • Agents
  • Tool use
  • Evals

Hiring an applied AI engineer in Boston (or remote ET)?

Email is the fastest channel. Résumé below.