AI engineer, open to full-time and contract.
I design, build, and ship production AI systems end to end: agents that do real work, retrieval that returns the right thing, evals that tell you whether it actually works, and LLM features that live inside real SaaS. I care as much about the parts that aren't glamorous (the eval harness, the failure analysis, the fallbacks) as the demo.
Below is the evidence, in the order I'd want you to weigh it: the code first, then shipped products, then the thinking behind them.
Read the code
The fastest way to evaluate me is to read what I've written. My flagship open-source artifact is FerroDB: a relational database built from scratch in Rust, now AI-native with an HNSW vector index. It's the “can you build the machine, not just glue one together” proof: every subsystem a production database hides, from page-based storage, B+-tree indexes, write-ahead logging with crash recovery, MVCC, and a cost-based optimizer, to enough of the Postgres wire protocol that real psql connects, built from first principles, covered by 122 tests, and documented in a live architecture book. Its ninth milestone adds an HNSW vector index with filtered k-NN and hand-written AVX2 SIMD kernels, the systems half of “AI-native.”
Closer to the AI work, text-to-sql-prompting: Measuring the accuracy lift of each prompting technique on text-to-SQL: execution-grounded evals, honest failure analysis. The readable AI-engineering code behind the positioning: a prompting ladder with an honest write-up of where it still fails.
More work lives on my GitHub profile.
Selected work
Three shipped projects that show the range: real LLM integration, a production security audit, and a live multi-tenant SaaS with daily users:
- Fetti: Money Tracker for Cash Workers, receipt OCR on Anthropic vision. Receipt OCR built on Anthropic vision: snap a photo, get a structured, draft expense entry. Live at fetti.tips.
- Justice Ledger: Legal Practice Management SaaS, dual Anthropic + OpenAI legal research. Dual Anthropic + OpenAI SDK integration for legal research and document summarization, plus a full security audit of 32 data-touching routes.
- PowderLedger: Ski & Snowboard Rental Management. Production multi-tenant SaaS in active daily use at a real rental shop: the depth of a system someone else can run and extend.
Writing
How I think about building with AI, the trade-offs, not the hype:
- Vector Index Internals: HNSW, IVF, DiskANN, and What Changes When You Write One in Rust
- Retrieval, Measured: Where RAG Actually Breaks
- Prompt Engineering Is Now a Subset of a Bigger Job
- ChatGPT vs Claude vs Gemini vs Copilot vs Meta AI: A Developer's Honest 2026 Comparison
- Your AI Agent Forgets and Degrades: Engineering Context Across Time
- AI Writes 1.7× More Bugs. Two Disciplines Decide Whether They Ship.
How I work on a team
I usually ship solo, but I work the way a team needs: clear specs, reviewable PRs, honest trade-off write-ups, and code someone else can pick up. I review my own work adversarially, and I document the decisions the AI couldn't make, so the next person inherits the reasoning, not just the diff.
Let's talk
If you're hiring for AI engineering work, full-time or contract, I'd like to hear about it.