Skip to content

SofaGenius

For AIML researchers and engineers, How many times have you kicked off a batch of training runs, then spent the next few hours refreshing dagster stages, W&B plots, hunting through GPU utilization, digging different conf

Code Video Cerebral Valley

Core trick

An agent calls monitoring APIs for W&B, Dagster, and GPU utilization, presenting a plan that the user approves before any job launches.

Tags

dev toolsagents

Description

For AIML researchers and engineers, How many times have you kicked off a batch of training runs, then spent the next few hours refreshing dagster stages, W&B plots, hunting through GPU utilization, digging different configs, SQL, data, and different base models? You're babysitting the GPU, instead of thinking about the research and real problem. Or you could lean back on the sofa and tell your AI assistant to do all of that for you. Sofa Genius turns the boring, procedural parts of ML research, monitoring, data inspection, scouting, launching jobs, into a conversation. You speak your intent, approve the plan, and the agent handles the grind. You can focus on the creative part.
lilyzhng/SofaGenius★ 121 forksPythonMITCreated Feb 15, 2026 · last push Apr 26, 2026

README excerpt

<p align="center">
  <img src="frontend/public/SofaGenius_v3.png" alt="SofaGenius" width="500" />
  <!-- <br><br>
  <strong style="font-size:1.5em;">Sofa Genius</strong>
  <br>
  <em>The research assistant that lets you stay on the sofa.</em> -->
</p>

---
> **Run your research like a CEO.** Your best ideas come when you're relaxed. Sofa Genius gives you an AI team that handles the grind — you stay in creative mode. [Built with Opus 4.6 Hackathon](https://cerebralvalley.ai/e/claude-code-hackathon/details)

## What Is This?

SofaGenius is two things:

### 1. ML Research Assistant (Web App)
For ML researchers and engineers — stop babysitting GPUs. Sofa Genius turns the boring, procedural parts of ML research (monitoring, data inspection, scouting, launching jobs) into a conversation. You speak your intent, approve the plan, and the agent handles the grind.

### 2. Multi-Agent Coordination Framework (Agent Org)
Run multiple AI coding agents as an organization — each with its own identity, role, and persistent memory. Agents coordinate through a file-based handoff protocol and Discord.

```
agents/
├── genius-growth/CLAUDE.md     # Genius Growth — content + tribe building
├── genius-builder/CLAUDE.md    # Genius Builder — ships code and tools
├── genius-researcher/CLAUDE.md # Genius Researcher — research, data, deep dives
├── handoff/                    # Agent coordination (status, specs, reports)
├── pr-rules.md                 # PR workflow rules
└── onboarding.md

Team