How I build
Anthropic stackI commit to a stack, then route by judgment.
Most AI shops say they are model-agnostic. In practice that means defaulting to whatever ran last, or whatever is cheapest the morning of the demo. I commit differently: to a stack, not a single API. Claude API runs primary reasoning and document work, Claude Code is my engineering layer, and MCP wires it into the tools a business already uses. Around that core I route deliberately — cheap classification on smaller models, embeddings on commodity infrastructure, and Gemini where it fits better, as it does for WhatsApp parsing in Denver Trades. Four production systems is not a huge sample, but it is enough to have made these calls for real rather than in the abstract.
The stack I commit to
- 01Claude CodeThe engineering substrate for every custom build.
- 02Claude APIPrimary reasoning, document work, and structured outputs.
- 03MCPConnectors built and maintained for the tools your team already uses.
Everything else
- Multi-model routing for cost-sensitive auxiliary tasks — GPT-4o-mini, Gemini Flash, open-source where it fits.
- A point of view on when not to use AI at all. Sometimes the answer is a spreadsheet, not a model.
