phnix · agent infra that survives the real run

Pure infra gives you products. Pure research gives you papers. The intersection gives you both, and that's the only place worth building.

the working thesis · 2026

blackreach 3k+ tests lethe behavioral drift huginn 800+ tests claude-voice ★25 rigr agent regression amp discovery 88.3% f1 local-first by default

selected work

public code · documented results

Research · public benchmark

Lethe

An open benchmark for how AI agents degrade over long runs. Not "can it do the task." The real question is whether it still behaves correctly after a hundred steps. Four metrics fold into a single Drift Severity score, and every outcome is verified by a real shell command, never the agent's own report.

0 · stable drifting agent → 8.0 SEVERE 10

A paper, an infra tool, and a question nobody else is benchmarking. The cleanest proof of the whole thesis.

read the methodology →

Public · 3,055 tests · video demo

Blackreach

A browser and research agent for real, stateful web work. It reduces noisy pages, preserves interrupted progress, and checks outcomes separately from the agent's own report. The flagship case study includes a cinematic 27-second real-web proof.

● ● ●

$ blackreach run "collect and verify sources"

✓ session resumed · observations updated

✓ results saved with source links

case study + demo → github ↗

Public · local-first · ★25

claude-voice

A fully local voice mode for AI coding sessions, with synchronized word highlighting. It runs without cloud APIs and is packaged for straightforward installation.

● ● ●

$ pip install claude-voice

✓ local speech ready · no API key

✓ word highlighting synchronized

Small enough to understand, useful enough that other developers installed and starred it.

view project → github ↗

Public · self-hosted

Huginn

A self-hosted web collection API for JavaScript-rendered pages and structured extraction. It is designed around explicit job state, retry behavior, and local operation without a per-page service fee.

● ● ●

$ huginn collect https://example.com

✓ rendered page · extracted content

✓ job complete · result stored

Public code, deployable locally, and backed by a test suite with more than 800 tests.

view project → github ↗

research & corpora

where infra meets the paper

AMP Discovery

Antimicrobial-peptide discovery by fine-tuning Meta's ESM-2 protein language model with LoRA. Built a leakage-aware protein screening experiment and kept the failed transfer result visible.

88.3% F1 (leakage-free)650M ESM-2 + LoRA

PyTorchESM-2LoRABio

read the writeup → github ↗

RLVR / GRPO Lab

An inspectable post-training harness with verifiable math rewards and strict answer-contract evals. v1 research run complete, with paired-bootstrap evidence behind every claim.

3B & 7B runsGSM8K, full split

RLVRGRPOEvals

read the writeup → github ↗

Library of Alexandria

A curated cross-cultural mythology corpus. Verified primary sources across 69 traditions, rebuilt from 172GB of junk down to ~67GB where every download is checked. A model-training substrate.

~675k curated files69 traditions

RAGCorpusCuration

view project →

02·5