The infrastructure layer for AI-native software — and the benchmarks to prove it.

Agents need three things production-grade: a runtime fast enough to serve them, a way to act on the world, and a way to keep their code honest. That's the lab: fast Python served by a Rust core, a Rust-first agent runtime, a browser agents can actually read, an AI reviewer that holds the line — each one measured, never just claimed. Don't say it's fast. Prove it.

A high-performance Python web framework powered by a Rust core. Built on Per-Interpreter GILs (PEP 684), it runs Python handlers across every CPU core in a single process — no per-process memory tax of multiprocessing.

• 902k req/s pipelined plaintext, 423k on the standard baseline (8C/16T).
• 2.7× Robyn at equal scale, in 1/3 the memory.
• Sustained 400k QPS: 4 MB RSS growth over 73.8M requests — ~0 B/req, zero leaks.

A Rust-first multi-agent LLM runtime. A dozen providers behind one trait, a composable middleware pipeline, an Agent abstraction you hand tasks to, and Python + Node bindings — with observability built in, not bolted on.

• Typed error hierarchy (Permanent / Retryable / RateLimited / Auth).
• Multi-tenancy enforced at every layer; cache hit/miss observable per call.
• Same Pipeline runs identically local (in-mem) and in a service (Redis + S3).

An MCP-compatible browser built for LLM agents. Where Playwright MCP dumps 25k tokens of raw HTML, SiliconSurfer returns finished data — clean Markdown, interactive element refs, structured JSON — across 5 vision modes.

• 30/30 extraction eval (vs Jina 20/30), 5/5 E2E (vs browser-use 0/5).
• 6.2× faster; ~1ms warm calls via a shared daemon.

A.R.C. (Adversarial Resolution Cycle) — a code-review tool where a Critic LLM files findings and a Fixer (Claude / Codex CLI) edits, looping until they converge or you step in. Single Rust binary; state in .arc/ (SQLite + SARIF). Free binaries (engine closed-source). An earlier open-source Python version lives at arc-cli.

These aren't four side projects — they're one stack. Pyronova serves agent workloads at multi-core speed; TARS is the runtime those agents run on; SiliconSurfer is how they read and act on the web; and arc keeps all of it honest by reviewing the code. Mostly Rust under the hood, every performance claim backed by a reproducible benchmark, built in public.

Writing soon at leocaolab-blog.pages.dev