◆ XeeNet

Distributed AI Research at Scale

An open platform for autonomous machine learning experimentation.
Donate spare compute. Run real PyTorch training. Advance open science.

Documentation · Architecture · Training Pipeline · Desktop Worker · Roadmap

---

What is XeeNet?

XeeNet is a distributed compute platform for ML research, inspired by SETI@home and Andrej Karpathy's autoresearch. Volunteers install a lightweight desktop app, donate idle CPU/GPU time, and the platform runs real neural network training across the global grid.

Researchers submit experiment campaigns. The orchestrator decomposes them into bounded, self-contained training tasks. Workers execute actual PyTorch training (not simulations), report metrics, and the dashboard aggregates results. A credits economy rewards contributors and funds research.

This is not a concept. The end-to-end pipeline is functional: research brief to real val_bpb metrics from distributed workers.

Key Features

• Real ML Training -- Character-level transformer on TinyShakespeare with genuine val_bpb metrics. Not simulated.
• Zero-Setup Desktop Worker -- Electron app auto-downloads Python 3.12 + PyTorch on first run. NVIDIA GPU auto-detection installs CUDA-enabled PyTorch. Users click "Start" and contribute compute.
• Autoresearch Contract -- Every task is self-contained, has a fixed time budget, and reports a single comparable metric. Dual-deadline pattern (soft at 90%, hard kill at budget + 15s) guarantees termination.
• Reproducible by Default -- Seeded config generation, deterministic training. Same seed, same config, same results.
• Multi-Agent Architecture -- Orchestrator, Worker, Portal, and Economics agents with defined roles and shared memory.
• Live Dashboard -- HTMX + Jinja2 web UI with real-time stats, campaign results, worker monitoring, and credit tracking.
• Credits Economy -- Workers earn credits for completed tasks. Researchers spend credits to run campaigns. Anti-fraud verification built in.
• Graceful Degradation -- No PyTorch? Workers fall back to simulated metrics with a clear UI indicator. The platform never blocks.

Screenshots

Dashboard Overview	Campaign Results
Platform stats with live worker count, task completion, and credits	10/10 tasks completed, best val_bpb 3.5705, full hyperparameter breakdown

Desktop Worker (Connected)	Desktop Worker (Training)
RTX 3080 detected, CUDA PyTorch ready, Real ML Training badge	Active training: spawning train_char_lm.py with hyperparameter config

See the full documentation site for detailed screenshots and architecture diagrams.

Architecture

Research Brief --> Orchestrator --> Task Queue --> Workers --> Training --> Metrics --> Dashboard

Component	Technology	Role
Backend API	FastAPI + async SQLAlchemy + SQLite	REST API, orchestration, persistence
Dashboard	HTMX + Jinja2 + Pico CSS	Real-time web UI, campaign management
Desktop Worker	Electron 28 + TypeScript	Cross-platform worker with auto-setup
Training	PyTorch (CPU or CUDA)	Real neural network training
Agents	Python (BaseAgent ABC)	Orchestrator, Worker, Portal, Economics
Config	Pydantic Settings + YAML	Type-safe settings with validation

How It Works

1. Researcher creates a brief defining the experiment goal, compute budget, and task count
2. Orchestrator decomposes the brief into tasks, each with a unique hyperparameter config sampled from a defined search space (learning rate, architecture, schedule)
3. Workers poll for tasks and execute them as isolated Python subprocesses with dual deadlines
4. Training script runs a character-level transformer on TinyShakespeare, outputs metrics as a single JSON line to stdout
5. Results aggregate on the dashboard showing best config, mean/std val_bpb, and per-task breakdown

Sample Results

From a real 10-task campaign:

Metric	Value
Tasks completed	10 / 10
Best val_bpb	3.5705
Std deviation	0.5622
Configs tested	10 distinct (varied lr, schedule, layers, heads, d_model)
Best config	lr=0.001248, step_decay, 2 layers, 4 heads, d_model=64
Wall time per task	~9s (CPU)

Quick Start

Server

pip install -e ".[dev]"
uvicorn services.api.main:app --reload
# Dashboard at http://localhost:8000/dashboard

Desktop Worker

cd desktop
npm install
npm run dev

Or download the portable .exe from releases. The app auto-downloads Python + PyTorch on first launch.

Training Script (standalone)

pip install torch
python experiments/train_char_lm.py --config-file config.json

Tests

pytest  # 110 tests, all passing

Project Structure

agents/              Agent definitions (orchestrator, worker, portal, economics)
skills/              Reusable modules (task generation, result analysis, scheduling, credits)
services/            FastAPI backend, database, schemas, orchestration logic
  api/               REST API routers + HTMX dashboard + Jinja2 templates
  db/                Async SQLAlchemy + SQLite
experiments/         Self-contained training scripts
desktop/             Electron desktop worker (TypeScript)
config/              Pydantic settings + YAML config
tests/               Test suite (110 tests)
static/              Dashboard CSS/JS assets

Contributing

Contributions welcome in any area:

• New experiment templates following the autoresearch contract (fixed budget, self-contained, single metric)
• Platform improvements to the dashboard, API, or worker capabilities
• Cross-platform testing of the desktop worker on diverse hardware
• Research ideas for what experiments the grid should run

License

MIT

Links

• Documentation -- Full platform writeup with architecture diagrams and screenshots
• Vision & Market -- Scale economics and the case for distributed autoresearch
• Roadmap -- Current state and future plans