From wiring buildings to wiring systems.

Decentralized creator tools where podcasters own their identity, feed, and audience. Built on the AT Protocol with custom lexicons and a Fastify/Next.js stack to keep creator data portable and under their control. Also published an AT Protocol skill on skills.sh to help others build on the protocol.

• 1 Federated ingestion at scale: resilient WebSocket consumer with bounded memory backpressure (1,000-event cap) and 100-event batching. Together with in-memory caching and longer cursor flush intervals, this cut production DB operations 96%, from ~209K to ~3K to 8K per day.
• 2 Audio extraction worker: dedicated ffmpeg worker with SSRF guards, FOR UPDATE SKIP LOCKED job claiming, and Cloudflare R2 content-addressed storage.
• 3 Multi-env production deploy: API on Fly.io, web on Vercel, Postgres on Prisma Postgres, with Sentry + Axiom + OpenTelemetry observability.

NOTE:I care about ownership and systems that don't trap people. This project pushes me to think beyond a single app and into protocols and ecosystems.

A production ELT pipeline and grounded LLM ask tool. Dagster orchestrates NFL play-by-play, schedules, and contract data (2015 to present) into Snowflake RAW, STAGING, and MARTS layers via dbt, then publishes finished marts to Postgres for a public website with a Claude-powered question tool that only states numbers it got from a query.

• 1 Production ELT: 20 Dagster assets with season partitions, refreshed on schedule by Dagster+ Serverless. Each load deletes, copies, and audits inside one Snowflake transaction and rolls back on failure; 43/43 dbt tests passing in prod.
• 2 Audited serving layer: finished marts publish to Postgres in a single-transaction swap, verified against Snowflake by row counts and primary-key fingerprints, so the website never queries the warehouse.
• 3 Grounded LLM Q&A: a Claude Haiku tool loop over select-only, parameterized, read-only queries with timeouts and row caps. The code rejects any number not returned by a query; an 81-question eval suite scores 78/81 (30/30 refusals) at ~$0.0005 and 4s per question.

NOTE:Ran like production: rate limits, spend caps, no stored user data, required CI checks, least-privilege roles, and a monthly resource monitor.

A real deployed app for saving, organizing, and analyzing travel. Combines a FastAPI/Postgres backend with a React frontend, including map views and basic analytics around visits and plans.

• 1 Full-stack architecture: FastAPI, PostgreSQL, and Docker with JWT auth, Pydantic validation, and a reusable CRUD service layer.
• 2 External API integration: city-based weather lookup with ~100ms response times, plus an interactive map layer for geographic visualization.
• 3 Calendar itinerary system: monthly/agenda views for trip planning with date-based organization of saved cities.

NOTE:It's a complete loop: database design, APIs, frontend integration, deployment, and iteration based on actual usage.

A pre-prompt middleware that condenses link-bearing Markdown by resolving remote links into compact cryptographic markers, cutting context token consumption and API latency for LLM workflows. Published on npm with zero runtime dependencies.

• 1 Context window optimization: condenses prompt payloads by replacing unchanged remote Markdown links with compact markers, drastically cutting token usage.
• 2 Conditional caching pipeline: two-phase network verification using HTTP HEAD probes and If-None-Match ETag validation to skip unchanged resources.
• 3 Zero-dependency architecture: CLI and library built exclusively on native Node.js APIs. No third-party supply-chain risk, minimal cold-start overhead.

NOTE:A focused, shippable tool that solves a real LLM-context problem with a craftsman's restraint: do one thing well, with no dependencies you don't need.