# Medha

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Dileep Kumar Salla](https://cerebralvalley.ai/u/DileepSalla)
- **GitHub:** https://github.com/DileepreddeS/medha
- **Demo video:** https://youtu.be/grqGhXAdkH0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3673

Problem: AI agents forget. Every task starts from zero, so the same chore costs the same tokens on day 100 as on day 1, and answers vary run to run. Letting agents write their own skills doesn’t fix this: 2026 research (SkillsBench) found self-generated skills give no benefit on average, because nobody verifies them.

Medha turns chores into verified, versioned skills: real code with a contract and tests.

Learn: the author model writes the skill.
Verify without self-grading: an isolated test writer that never sees the code, hidden holdout tests to catch overfitting, and a different model (Muse Spark 1.3) as judge.
Approve and run safely: a human approves new skills, which execute in network-blocked Modal sandboxes.
Reuse: known tasks skip authoring, testing and judging.
Repair: when inputs drift, Medha revises the skill, and the new version must pass every past case before going live.
Compose: it combines existing skills into larger workflows.

Who it helps: non-programmers with recurring digital chores (small business owners, ops staff, students) who get reliable automations without writing code.

Results (live benchmark runs): on repeated tasks, 2.6× fewer tokens and 3.3× faster than a single-shot agent; identical output 100% of the time vs 20%; drift repairs at less than half the cost. A single-shot agent passes slightly more tasks on first try (11/12 vs 10/12) but pays full price every time.

Meta stack: built with Muse Code; Muse Spark 1.3 as the judge; Muse Glimmer (open weights) self-hosted on Modal with vLLM as the author option (7/12 so far, with failures caused by cold-start timeouts). The demo video replays recorded model responses to fit 3 minutes; skills execute live.

## More from Meta Global AI Developer Hackathon

- [prefixgate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3670)
- [Zazen Builder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3671)
- [Italian Rides](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3672)
- [Agentztein](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3674)
- [Team Juicy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3675)
- [DeepKomal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3676)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3673. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt