How the work actually shows up.
Cost controls, evaluation suite, prompts under version control. An AI project without an eval pipeline isn't a project — it's a demo.
01
Anthropic-first integrations — caching, cost caps, structured outputs where they earn their keep.
02
Evaluation suites and prompt/version discipline before anything touches users.
03
MCP servers when tool reuse across clients or agents justifies the surface area.
04
Structured extraction into Postgres — typed records, not mystery JSON blobs.
Deliverables
- Prompt engineering + evaluation suite
- Anthropic / OpenAI API integration with caching + cost controls
- MCP server(s) for tool orchestration
- Structured extraction pipeline (Postgres ingest)
Stack
- Anthropic Claude (API + MCP)
- OpenAI (where justified)
- Python / FastMCP
- TypeScript
- Postgres
- Sentry
Real work, no stock.
Mockups and production screens from client and own-product work across the ai & automation discipline.
Built with this discipline.
Anthropic (Claude) is my default for reasoning + tool use. OpenAI where a specific capability wins. Local models only when residency or cost forces it.
Pick
a week.
I'll have it live
the one after.
No slot fits? Send a brief →