I build production AI systems and open-source the tooling around them.
Principal Engineer at Squiz · Brisbane
At Squiz I built the AI infrastructure from zero: production RAG on AWS and Cloudflare, ingestion across 1M+ pages, and an S3 Vectors integration AWS featured at re:Invent 2025. Below are the systems and tools I build and ship in the open.
- games-games-games → games.jasonmatthew.dev: a RAG rules lawyer for tabletop games. Hybrid retrieval (BM25 + reciprocal rank fusion), cross-encoder reranking, cited answers, and an eval harness that gates every retrieval change.
- formrecap-lora-classifier → lab.formrecap.com: 7 LoRA adapters across 4 model families, benchmarked head to head. Best F1 = 0.961 (Mistral 7B). Dual-deployed: Modal vLLM (calibrated) + Cloudflare Workers AI (edge).
- claude-skills: my Claude Code plugin marketplace. 9 plugins for orchestration guards, tiered deep research, subagent-driven development, ADRs, session retros, handoffs, and cross-provider adversarial review.
- skopia → skopia.dev: privacy-first web analytics you self-host on Cloudflare's free tier. No cookies (visitor identity is a daily-rotating HMAC), exact rollups from a Durable Object, 554-byte tracker. AGPL.
- formrecap → formrecap.com: form abandonment recovery SaaS. ~3KB snippet, 100% Cloudflare-native (D1, KV, Workers, Queues, Durable Objects). Per-customer AES-GCM encryption. 225 automated tests.
- pipespy: real-time pipeline debugger for the terminal.
pvshows bytes, pipespy shows your data. Distributed via homebrew tap.
- Long-form: jasonmatthew.dev
- AI engineering and leadership posts: LinkedIn
- Short-form: @jasonm4130




