Phillip and Claude
performance improvements for compiled binaries are not always obvious or easy to do without claude figuring out and redoing a ton of work each time costing tokens.
Core trick
Profiles hot paths, feeds the LLM a token-budgeted source-attributed brief, and commits only changes that pass tests and run faster.
Description
performance improvements for compiled binaries are not always obvious or easy to do without claude figuring out and redoing a ton of work each time costing tokens. Make part of the profiling and benchmarking deterministic so that Claude can have an easier and more accurate time attempting to find performance improvements
README excerpt
# autonomous-profiler
**An autonomous performance engineer for Rust.** Point it at a codebase; it profiles the
real hot paths, hands an LLM a *token-budgeted, source-attributed* brief of where the time
goes, then runs a closed loop that optimizes the code and **commits only changes that are
proven correct and measurably faster** — with a live dashboard of every attempt.
> 🔴 **Live dashboard:** https://claude.autoperf.run · 🏆 **[Highlight: it made polars 4% faster, provably](highlights/polars-resolve-chunked-idx.md)**
---
## What it actually did (results)
| Target | Outcome | Gate |
|---|---|---|
| **analyzer** (internal polars/ndarray analytics lib) | **−39%** flights / **−16%** transactions, **7 commits** | full `cargo test` suite |
| **polars** (the OSS dataframe library itself) | **−4.0%** on the sort/gather path, **1 commit** ([details](highlights/polars-resolve-chunked-idx.md)) | correctness fingerprint over 2.96M rows |
| swarm tournaments | diverse-lens rewrites; the **algorithmic lens landed a quantile-calculator win** the serial loop missed, while slower candidates self-reverted | benchmark + correctness |
The reverts matter as much as the wins: the loop **refuses to commit a regression or an
unverified change**. It also tried to out-hand-tune OpenBLAS/SIMD and correctly failed —
no fake wins.
---
## How it works (practical walkthrough)
```
1. ap run <target> # profile with samply, symbolicate, rank hot functions
│ # → token-