# RAISEMYKNOWLEDGE - **Event:** [RAISE Summit Hackathon](https://cerebralvalley.ai/e/raise-summit-hackathon) - **When:** Jul 4 at 9:00 AM – Jul 5 at 7:00 PM (GMT+2) - **Where:** Paris, France - **Team:** [Jose Contreras](https://cerebralvalley.ai/u/Pepe), [Rares M](https://cerebralvalley.ai/u/Rares), [Ceco Simeonov](https://cerebralvalley.ai/u/Ceco), [Nicholas Palade](https://cerebralvalley.ai/u/nickpalade) - **GitHub:** https://github.com/nickpalade/Brujula - **Demo video:** https://youtu.be/8HsCqRyrP3U?is=VRdBeMqsAYecnUBB - **Gallery:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/70 Built in the spirit of the Google DeepMind track mission: build AI efficiently to benefit humanity. On June 25, 2026, twin earthquaks (M7.2 + M7.5) struck northwestern Venezuela. Comms and Power failed. 50000 dissaparances. Dozens of Agencies, volunteers and survivors were all reporting at once: by voice, text and photo in spanish often messy and contradictory. Climate change is making natural disasters more common and stronger, there is a coordination failure: duplicate teams sent to one collapse while another site waits, water trucks idle near shelter with no drinking water, missing person names scattered across raadios and chat threaads. Responders fall back to white boards because the digital tools need connectivity they do not have and cannot risk private information of people being scattered. In the disaster zone the network is done or oversubscribed. Cloud AI is unreachable. Compass is built on exactly this premise: No internet what so ever, only a laptop hostpot and phones on local WI FI. How we solve this problem: 1 Field phones submit reports — voice, text, photo, GPS when available — through a store-and-forward outbox that survives radio drops. 2 Gemma on the command-post laptop runs a multi-step agent workflow: parse → dedup → prioritize → match. Retrieval (protocol KB, open incidents, resource inventory) is one tool inside that workflow — not the product. 3 The coordinator confirms every dispatch. Gemma proposes; humans decide. Override, edit, rematch, and broadcast alerts stay in human hands. 4 Assignments sync back to field phones over the LAN — crews and volunteers get tasking to go to the place: incident location, summary, and why they were sent. They accept the mission, mark En camino → En el sitio, and report when done. Late-registering volunteers trigger new match proposals for open needs. The command graph makes this visible: field reports flow into a central Gemma brain node, structured incidents and dispatches connect out. The agent's reasoning is the product; feed, map, and graph are windows onto it. Privacy by design: Reports contian missing people identities, medical status and locations of vulnerable peoplpe. In a politically charged setting routing that through a foreign cloud is unacceptable. Now the technical part, this is a summary, please refer to the README for the full specs. Architecture. One Express server (:8000) on the command-post laptop hosts everything: API, the React app (Command Post + field client), SQLite store, offline map tiles, and an embedded Ollama child process running gemma3n:e4b. Phones install the field PWA from a QR code over the laptop's hotspot — hub-and-spoke, no app store, no cloud account. Bootstrap once with internet, deploy offline indefinitely. Same codebase renders both surfaces from shared design system and API contracts. Reliability engineering. Every Gemma step uses Ollama structured output with a JSON schema, then server-side Zod re-validation; malformed output retries once, then falls back — the hub never 5xxs a field phone. Dedup and match IDs are enum-locked to real board entities, so the model cannot hallucinate an incident or resource that doesn't exist, with a category-compatibility backstop on top (rescue ↔ machinery merges allowed; unrelated categories blocked even if Gemma guesses wrong). Prioritization is fully deterministic (urgency × people × age, zero model calls — fast and always explainable). If Gemma fails mid-pipeline, the report is stored as pending and the board keeps working; if the knowledge service is down, protocols fall back to a local KB. Reports are acked in ≤20 s with slow parses finishing in background. Sync & field resilience. Delta polling (GET /api/sync?since=) with monotonic sequence numbers; idempotent client_ref prevents duplicates on flaky reconnects. Voice goes phone → hub for local STT (audio never stored); photos are compressed on-device and passed to the parse step only — the hub stores has_image: true, never the bytes. Reports work in 20+ languages with board summaries in a configured output language. Location resolution: phone GPS → offline gazetteer → no pin (incident still boarded). Beyond the core loop. One-click SITREP generation, escalation watchdog (unattended critical/high incidents auto-flagged), missing-persons registry extracted across incidents, broadcast alerts, human-correction tracking (corrected_by_human badges), full dispatch lifecycle with outcome notes, GPU/CPU toggle and in-browser model manager, and LAN access control (Command Post restricted to the coordinator station). Verification & compliance. npm run verify:hub replays the full demo flow in 15 end-to-end checks; 18 model-free contract tests plus Playwright UI tests for the graph. Fully open source (Apache-2.0 Gemma), built entirely within the hackathon window (Sat 11:30 → Sun 12:00), and outside the banned categories: not basic RAG, not Streamlit, not a medical-advice bot, not a dashboard-as-product — advisory content is sourced humanitarian protocol (INSARAG, START, Sphere, PAHO) for trained responders, and nothing dispatches without human confirmation. ## More from RAISE Summit Hackathon - [low cortisol - FlowTwin](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/67) - [AI Collective Kansas City](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/68) - [PRAETOR](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/69) - [Overturn](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/71) - [Hello](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/72) - [Flashdeal](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/73) --- Markdown version of https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/70. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt