# Aloo Paratha

- **Event:** [Google DeepMind Bangalore Hackathon](https://cerebralvalley.ai/e/google-deepmind-bangalore-hackathon)
- **When:** Sat, Jul 11 at 9:00 AM – 10:00 PM (GMT+5:30)
- **Where:** Marathahalli, Marathahalli Main Road
- **Team:** [Kunal Achintya Reddy Seerapu](https://cerebralvalley.ai/u/kunalachintya)
- **GitHub:** https://github.com/latent-0/mao-arch
- **Demo video:** https://youtu.be/enuSFdxFV-Y
- **Gallery:** https://cerebralvalley.ai/e/google-deepmind-bangalore-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/google-deepmind-bangalore-hackathon/hackathon/gallery/31

Multi-agent handoffs today rely on flat prose summaries. This lossy projection discards structural graph dependencies and resource constraints, forcing receiving agents to guess step order and rules. This causes a 51.3% failure rate under task-specification constraints. Appending growing conversational histories also causes severe KV-cache pollution and context bloat.

The Solution (Mao)
Mao is a plug-and-play middleware SDK that replaces prose handoffs with a joint structural-semantic vector space (128-dimensional). It projects a relation-aware graph attention network (R-GAT) representation of the active task frontier and aligns it contrastively with agent reasoning traces using a triplet margin loss. This clusters valid actions close to the graph state while pushing violating actions far away.

Local Safety & Witness-Routing
At runtime, Mao measures the cosine similarity between the graph and the proposed action. For safety, it uses a hybrid neuro-symbolic gate called witness-routing. An action is only rejected if the neural gate fires and the symbolic graph checker can name the specific violated constraint edge. This ensures a flawless gating precision of 1.000. On rejection, local Gemma 4 (via Ollama) translates raw topological violations into natural-language replan instructions.

Key Metrics (n=300 held-out samples)
Baseline Prose Handoff Success: 48.7%
Mao Local Offline Success: 99.7% with a median latency of 1.7 ms.
Mao Gemini Cloud Success: 99.3% with a median latency of 527 ms.
Zero-Shot Domain Generalization (LOTO): Averages 95.1% success on unseen workflows.
Out-of-Distribution Safety: Deferral rate safely climbs from 6% to 62% under domain shift rather than failing silently, keeping precision at 1.000.000 precision everywhere.

---

Markdown version of https://cerebralvalley.ai/e/google-deepmind-bangalore-hackathon/hackathon/gallery/31. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt