I build the infrastructure layer
for coding agents
- § 01 Models are commoditizing. The harness is where the craft lives.
- § 02 70% of agent failures are context errors, not model errors.
- § 03 The best tools quietly raise the ceiling.
- § 04 Classical ML is underused on the outside of the model.
- § 05 Production-first — theory only matters if it ships.
- § 06 Every sentence earns its place. All signal, no noise.
What I'm working on.
At Strange Loop Labs — a forward-deployed AI shop for Fortune 500 and Big Four clients, built by Amazon Alexa engineers — I own technical architecture and delivery across client engagements: exec discovery through production, regulated enterprise document workflows at scale.
Internally I'm researching high-throughput agentic coding — applying classical ML (ensemble methods, online learning, statistical quality control) to the code-generation loop, not just prompt engineering. Map-reduce-style ticket-driven development with ticket-rs as the context graph.
On the side I'm building the tools I wanted to exist: ticket-rs.io for context graphs coding agents can actually use, sgrep.sh for sub-10ms semantic search in pure Rust, and SQLGenie for natural-language data access at enterprise scale.
Context is the new prompt. 70% of agent failures are context errors, not model errors — and the scaffolding around the model is a 10–100× quality multiplier.
I write Vanishing Gradients from the forward-deployed trenches. No theory without receipts; no claim without code.
Tools I'm building.
- Public alpha · 2026ticket-rsAI-native issue tracking. Git-backed, local-first, PageRank-prioritized. MCP servers for Claude Code, Cursor, Codex, Copilot, Cline, Gemini, Windsurf, Zed, JetBrains. MIT-licensed.Coding Agents
- Public alpha · 2026sgrepGrep by meaning, not keywords. 8M-parameter Model2Vec via Candle, pure Rust, sub-10ms semantic search. One static binary, zero API calls. MIT-licensed.Developer Tools
- Founder & CTO · 2024SQLGenieNatural-language query engine across 23+ SQL dialects. ~200× lower per-query cost than the manual baseline. Inbound from TechStars, multiple VCs, and two acquirers in the first 90 days.Data
Selected writing.
All writing (opens in new tab)- Level Up Coding3× faster file conversion with DuckDB.Data engineering
- Level Up CodingThe modern data stack is dead.Thesis
- Level Up CodingDipping your toes into JavaScript.Cross-language
- AI in Plain EnglishWhat the heck is a classifier?Primer
- AI in Plain EnglishDecision boundaries.ML foundations
- AI in Plain EnglishOn text similarity search.Retrieval
- AI in Plain EnglishDetect anything in text — GLiNER for zero-shot NER.NER
- AI in Plain EnglishEfficient text classification with GLiClass.Classification
- AI in Plain EnglishThe Qwen 2.5 model family.LLM survey
- AI in Plain EnglishGPT anywhere — running LLMs locally with Ollama.Local LLMs
- AI in Plain EnglishFour cool AI developer tools I've been using lately.Tooling
The arc.
Economics → data science → machine-learning science → applied AI → forward-deployed engineering. Each transition sharpened the same edge: pattern recognition across markets and systems.
Before Strange Loop Labs I was Principal AI Scientist at a stealth AR/VR startup, shipping a real-time manufacturing copilot on Meta Quest 3 and Apple Vision Pro. Before that: I owned the size-recommendation engine at True Fit — the company's flagship product, serving millions of shoppers daily across 30K+ brands including Nike, Target, and Walmart.
MS in Data Science, BA in Economics — CSU Fullerton. Writer for AI in Plain English and Level Up Coding. Topmate mentor.