latentra

Turn college campuses into AI supercomputers 💡

Inspiration

It's 2 AM. A grad student is waiting for GPU access. Position 34 in queue. Wait time: 6 days. Meanwhile, 10,000+ computers sit idle across campus. That's enough compute to run GPT-4 level models. But nobody can access it. We built latentra to fix this. 🎯

What it does

latentra is a peer-to-peer AI agent platform that pools idle campus computers to run massive AI models with powerful tool-calling capabilities. Students: Build and run AI agents without expensive GPUs or cloud bills Researchers: Deploy complex multi-step agents that use real tools (APIs, databases, scripts) Universities: Cut compute costs by 60-80% Planet: 80% less energy vs cloud computing (MIT research) How it works: Install desktop app on your laptop When idle, your device joins the campus compute pool Build AI agents with drag-and-drop interface Agents run distributed across campus using LocalAI's P2P inference Composio enables agents to call 100+ tools (GitHub, Slack, Gmail, etc.) Chroma vector DB stores agent memory across the network Contribute compute → earn credits → use for research 🛠️

How we built it

Tech Stack: Electron - Cross-platform desktop app React + TypeScript - Agent builder UI LocalAI - P2P distributed inference across campus devices Composio - Tool calling integration (100+ APIs and tools) Chroma - Distributed vector database for agent memory Python + PyTorch - Inference engine libp2p + WebRTC - P2P networking layer OAuth 2.0 - Campus SSO authentication Architecture: Key innovations: Agent builder with P2P inference: Visual interface to create agents that run distributed across campus Composio integration: Agents can actually DO things (send emails, query databases, execute code) Distributed memory with Chroma: Agent context and embeddings stored across network nodes Fault tolerance: Auto-recovers when students close laptops mid-agent-execution Campus-aware routing: Optimizes for network topology (same building = faster) Security: Sandboxed tool execution, encrypted everything, IT admin controls 🚧 Challenges P2P + Vector DB = Hard - Syncing Chroma embeddings across distributed nodes while maintaining consistency. Built custom replication protocol. P2P + Vector DB = Hard - Syncing Chroma embeddings across distributed nodes while maintaining consistency. Built custom replication protocol. Tool calling security - Students could potentially execute dangerous API calls. Implemented permission system and sandboxing with Composio. Tool calling security - Students could potentially execute dangerous API calls. Implemented permission system and sandboxing with Composio. LocalAI coordination - Managing inference across heterogeneous devices (different specs, availability). Built adaptive scheduling. LocalAI coordination - Managing inference across heterogeneous devices (different specs, availability). Built adaptive scheduling. Agent state management - When a node drops mid-execution, agent needs to resume. Added checkpointing at every tool call. Agent state management - When a node drops mid-execution, agent needs to resume. Added checkpointing at every tool call. Network instability - WiFi drops, laptops close. Built aggressive fault tolerance and state recovery. Network instability - WiFi drops, laptops close. Built aggressive fault tolerance and state recovery. 🏆 Accomplishments ✅ Built working AI agent that uses Composio to query GitHub, analyze code, and send Slack notifications - all running distributed ✅ Chroma vector DB successfully replicated across 10+ campus laptops ✅ Agent survived 3 laptops disconnecting mid-execution (fault tolerance works!) ✅ 240ms average inference latency with LocalAI P2P ✅ 73% energy savings vs cloud (measured with actual wattmeters) ✅ Successfully demonstrated agent making real API calls through Composio ✅ UC Berkeley Research IT interested in piloting 📚

What we learned

Technical: LocalAI's P2P capabilities are powerful but need careful orchestration Composio's tool integration is amazing - 100+ tools out of the box Distributed vector databases (Chroma) need smart replication strategies Agent state management across unreliable networks is genuinely hard Product: Researchers want AGENTS, not just inference - they need tools to interact with their workflow The combination of P2P inference + tool calling + memory is uniquely powerful Students love the visual agent builder - makes AI accessible Process: Integration is key - LocalAI + Composio + Chroma work beautifully together Demo with real agents doing real tasks is WAY more impressive than just running models Building for campus first (not general cloud) gives us unique advantages 🌟 Impact If adopted at UC Berkeley: 💰 Financial $200-2,000 saved per student researcher/semester $1-5M campus-wide annual savings Free access to agent tools that normally cost $50/month per user 🎓 Academic Every student can build AI agents with real capabilities Research automation: agents that run experiments, analyze data, write reports Democratizes AI research access 🌍 Environmental 450 tons CO₂ saved annually 80% energy reduction vs cloud Uses energy already being consumed (idle devices) 🤖 Agent Capabilities Students can build research assistants that actually interact with GitHub, Notion, Slack Multi-step workflows automated across distributed compute Persistent memory through Chroma for long-running research projects 🚀

What's next

Short term: Pilot at UC Berkeley with 3 research labs Deploy to 5 computer labs (2,000+ devices) Expand Composio integration to more research tools (LaTeX, Jupyter, ArXiv) Medium term: Expand to all 10 UC campuses (280k students) Multi-campus agent federation (agents that span Berkeley + Stanford networks) Add agent marketplace - students share and remix agents Long term: 50+ universities nationwide Open source the agent builder and P2P orchestration layer Work with NSF to standardize distributed AI agent infrastructure Technical roadmap: More Chroma optimization for faster retrieval across P2P network Composio enterprise tools (databases, cloud platforms) Agent-to-agent communication for multi-agent systems LocalAI model switching based on task complexity Built at Cal Hacks 2025 Democratizing AI agents. One campus at a time. 🤖🌍 🔧 Tech Details Why these tools? LocalAI: Enables P2P inference without centralized servers - perfect for campus networks Composio: Gives agents real-world capabilities - 100+ integrations out of the box Chroma: Lightweight vector DB that works distributed - agent memory across devices What makes this powerful: Traditional cloud agent platforms = expensive + centralized + energy-hungry latentra = free + distributed + sustainable + runs on campus hardware you already have DEMO VIDEO : link