AI Assist
- Pair-programming: Claude Code (CLI) + Codex (in-IDE) running in tandem.
- Web companion: Claude.ai for longer-form thinking; Anthropic API for build-time tasks.
- Frontier reach-for: a mix of frontier models situation-by-situation; the test harnesses (see SUM's multi-LLM cross-corpus benches) keep us honest about which one is best for what.
Dev infrastructure
- Hosting: Cloudflare Workers + Pages for production. Custom domains routed through Cloudflare DNS.
- Databases: PostgreSQL (via Supabase / Neon as appropriate), SQLite for embedded use, ClickHouse for OLAP.
- Vector / search: Qdrant, pgvector, depending on the project shape.
- CI/CD: GitHub Actions, Cloudflare Pages auto-deploy.
- Shell: zsh.
Languages
- Daily drivers: TypeScript, Python.
- For systems work: Rust, Zig, occasional Go.
- Frontend frameworks: Astro, SvelteKit, Next.js, React.
Process
- Spec-driven discipline where the project warrants it (see SUM and HFAO's PROOF_BOUNDARY.md / SPEC.md patterns).
- Outcome-not-mechanism portfolio framing for monetization-track projects; mechanism stays in private repos and case studies.
- Verifiable claims over marketing language.
Personal hardware and editor specifics aren't published — they're tools, not product decisions, and they change. The infrastructure, languages, and process above are what actually shape what ships.