Crisis Averted

Close the humanitarian funding gap with our Crisis Topography Command Center. We map aid mismatches on a 3D globe. Ask Pablo, our AI voice agent, to fly you to any country for deep insights.

Comment 1

Inspiration

We were inspired by the Hacklytics 2026 challenge on inequality in humanitarian funding allocation. Currently, billions of dollars in aid are spent reactively, driven by media visibility and sentiment rather than by systemic, measured need. We wanted to reframe global humanitarian aid as a quantitative capital allocation challenge. Our goal was to build a visceral, agentic interface, a 3D command center, where funders and responders could identify, analyze, and predict humanitarian crises through a living topography of data, commanded simply by voice.

What it does

Crisis Averted is a 3D geospatial intelligence platform that maps humanitarian funding mismatches in real time. It features a custom-engineered 3D environment where data is physically manifested as interactive geometry via our Topographic Mismatch Engine, which calculates and projects the Severity-to-Funding Gap Ratio as 3D extrusions on the globe.

The core of the experience is Pablo, the sovereign AI orchestrator of our Command Center. Commanded solely through natural language, Pablo bridges human intuition and machine intelligence. Utilizing high-fidelity WebRTC streams, you can tell him, “Take me to the border of Sudan,” and he will physically fly the 3D globe to the precise location, adjust atmospheric glows to highlight semantic anomalies, and cross-reference analytical clusters to deliver personalized intelligence briefings.

The Distributed Agentic Pipeline (The "Engine")

We didn't just build a wrapper; we architected a multi-stage, event-driven ecosystem where voice, vision, and vector intelligence converge. We call this our Seven-Phase Resonance Loop:

  1. Acoustic Isolation & Ingress: High-fidelity voice input is captured via WebSockets/WebRTC and passed through the ElevenLabs Voice Isolator. Noise is stripped before hitting our primary LLM orchestrator.
  2. Agentic Dispatch: Gemini 3.0 Pro acts as the "Central Nervous System," performing real-time function calling to decompose natural language into executable system commands.
  3. Command Bus: Commands are published to a high-concurrency Kafka Stream. This ensures sub-millisecond dispatching to client-side WebGL triggers and backend analytical engines.
  4. Multi-Model Ensemble Execution:
    • Analytical Reasoning: Complex SQL queries are generated and executed against Databricks Genie Spaces for deep structured data mining (powered by Llama/Gemini).
    • Semantic Retrieval: High-dimensional embeddings of 18,000+ humanitarian projects are searched within the Actian Vector DB using actiancortex.
  5. Asynchronous Aggregation: Results from Databricks and Actian are fed back into a secondary Kafka Stream, serving as a global state synchronizer for both the 3D frontend and the narrative generator.
  6. Narrative Synthesis: The raw data (anomaly scores, ROI deltas, crisis metrics) is synthesized into a human-readable "Intelligence Briefing" by Gemini 3.0 Flash.
  7. Voice Egress & Haptic Feedback: The briefing is streamed back via ElevenLabs, while the 3D globe physically deforms its geometry to represent the unmet funding gaps.

How we built it

  • The Visual Layer: Built with Next.js 14 and react-globe.gl (Three.js/WebGL) to create the interactive 3D client.
  • The Intelligence Core: Real humanitarian data (UN OCHA/HDX HAPI v2) is ingested into a Databricks Lakehouse using Delta tables. We utilize sentence-transformers/all-mpnet-base-v2 to create dense vector representations of geopolitical narratives.
  • Infrastructure: To handle the high-concurrency demands, we orchestrated a dedicated Actian Vector DB across 192 Core CPU Vultr High-Compute VM. Unlike shared cloud databases, this self-hosted Actian instance on Vultr allowed for granular performance tuning, ensuring sub-100ms vector similarity searches so heavy query loads didn't bottleneck our voice agent. FastAPI handles the routing.

Challenges we ran into

Translating continuous audio streams into reliable, deterministic client-side UI commands was incredibly complex. We had to design strict tool contracts so Pablo could seamlessly control the 3D globe's state and camera without hallucinating invalid coordinates or desyncing the UI.

Integrating Databricks presented a steep learning curve, we had to optimize caching, manage cold starts, and work within table limits to ensure the backend didn't block the agent. Additionally, getting the massive Vultr cluster setup working and understanding how the 192 VMs fit into our vector DB architecture took intense effort, but it was absolutely essential for making semantic search fast enough for a real-time conversational UX.

Accomplishments that we're proud of

We are immensely proud of constructing an end-to-end, multi-model agentic pipeline from scratch. Successfully orchestrating WebRTC voice ingress, Gemini function calling, Kafka event streams, Databricks analytics, and Actian vector retrieval into a cohesive, sub-second feedback loop is a massive technical feat. We explored a ton of new ground and managed to turn a cold data dashboard into a living, breathing tactical partner.

What we learned

We gained deep, hands-on experience with modern data orchestration and high-performance computing. We learned how to model real-world humanitarian data in Databricks, generate dense vector embeddings, and run semantic search at massive scale with Actian on Vultr. Furthermore, we mastered the nuances of building an agentic UI, specifically, how to keep a complex 3D WebGL state perfectly synchronized with the asynchronous, streaming outputs of a conversational AI.

What's next for Crisis Averted

The map is not the territory. But a predictive, intelligent map can save the territory before it burns. Our ultimate vision is to deploy Crisis Averted as a live operational tool for organizations like the UN. We plan to harden the distributed pipeline and expand our predictive risk models, leveraging vector anomalies to forecast systemic risks like mass migration or famine before they escalate into catastrophes. We want to permanently reframe how global aid is distributed, treating it as a portfolio optimization problem to ensure every dollar yields the highest impact.

Built With

Share this project:

Updates