AgentHN

AgentHN (Agent + Hypernetwork): agents that write to their own weights at inference time, built on Doc-to-LoRA.

Tracks live in their own folders under src/agenthn/ so the team can work in parallel:

Folder	Owner	What
core/	shared	D2L model wrapper + config (keep stable)
personalization/	Eric	running profile doc → per-user adapter
memory/	Bryan, Nikash	long-horizon memory via stacked adapters
skills/	David	self-improving skills (T2L)
webapp/	Eric	demo UI

Setup

This package reuses the working doc-to-lora venv (which already has the full GPU stack + the ctx_to_lora editable install). It is installed editable into that venv — no separate environment needed.

export PATH="$HOME/.local/bin:$PATH"
uv pip install -e /home/ubuntu/itc-hackathon --python /home/ubuntu/doc-to-lora/.venv

Requires HF login for the gated google/gemma-2-2b-it base model (uv run --no-sync huggingface-cli login in the doc-to-lora repo).

Smoke test

/home/ubuntu/doc-to-lora/.venv/bin/python scripts/smoke_test.py

Web app (cache-only demo UI)

The deployed browser loads only committed files from webapp/static/fixtures. It never health-checks or sends requests to the Prime Intellect/GPU backend. Captured streams replay one frame every 200 ms; personalization and skills use recorded example choices so every enabled action has a genuine cached result.

The FastAPI/GPU code remains available only for regenerating fixtures. See scripts/README_GPU.md, including the command for capturing Large memory runs.

Layout

src/agenthn/
  core/
    config.py               # paths (D2L repo, checkpoint), device
    model.py                # D2LModel: load / internalize / snapshot / restore / chat
  personalization/
    extractor.py            # turns -> {category, value, action} updates
    profile_store.py        # per-user profile docs + cached adapters (swap)
  webapp/
    app.py                  # FastAPI: serves the SPA + personalization API
    service.py              # live (D2L) + mock services behind the API
    static/                 # index.html, app.js, styles.css (the demo page)
  memory/ skills/           # teammates' tracks
scripts/smoke_test.py       # load checkpoint, internalize, generate