First prod-traffic latency dashboard for LLMs. Real telemetry from f3dx + llmkit users who opt in. Cloudflare Worker ingest, R2 parquet, Evidence.dev + duckdb-wasm dashboard.

profile readme

Know what your AI agents cost. API gateway with budget enforcement, session tracking, and MCP tools.

Security linter for open source projects. Finds and fixes what others only report.

Calibration metrics for pydantic-evals: ECE, MCE, ACE, Brier, reliability diagrams, Murphy 1973 decomposition, temperature/Platt/isotonic scaling, Fisher-Rao geometry. The calibration layer the eva…

In-process Rust router for LLM providers. Hedged-parallel + 429/5xx hot-swap < 1ms. Composes with hosted gateways like llmkit.

Content-addressable LLM response cache + replay layer. redb + RFC 8785 JCS + BLAKE3. Drop-in for f3dx, openai, anthropic. Zero-cost CI runs, deterministic regression tests on prod traces.

Replay-driven eval for f3dx and pydantic-ai traces. Take a JSONL trace, swap the model, get a per-case diff.

f3d1 Rust-core runtime for pydantic-ai. Polars for agents. Drop-in for openai + anthropic SDKs with 5x streaming. AgentRuntime with concurrent tool dispatch (5-10x). Native Logfire/OTel emission.

The keystone of the f3d1 arch. Rust-native concurrent agent execution runtime. Phase 0 architectural spike passed; 24-week build greenlit.

Org-level community defaults for smigolsmigol/* repos. Inherited by every public repo without an override.