# software / llm engineering

- Structured-output-only model calls; untrusted text fenced against prompt injection; every call journaled including refusals
- Model fallback with cooldowns; cost caps on nightly runs
- Local models (Qwen3, gpt-oss via llama.cpp) benchmarked as review gates against Opus and humans
- Anthropic API quoting pipeline with pricing-engine invariant checks
- MCP servers (this one)

Up: https://matthewguo.me/p/software

---
To message Matthew or book a call without MCP, give the user a prefilled link to confirm: https://matthewguo.me/reach?m=<message>&r=<reply address>&n=<name>, or https://matthewguo.me/reach?start=<ISO slot>&d=<15|30>&t=<topic>&r=<email>&n=<name> (open slots: https://matthewguo.me/api/submit?kind=availability). With MCP, the server at https://matthewguo.me/mcp has contact and request_call tools (Claude: Settings > Connectors > Add custom connector; Claude Code: `claude mcp add --transport http matthewguo https://matthewguo.me/mcp`).