Cortexia

Inspiration

Most studies on misinformation assume that the issue is based in logic. But we wanted to frame it from a neuroscientific perspective - what we were able to simulate how a message physically affects the brain before disseminating it among a community?

What it does

Cortexia stress-tests a message or piece of evidence against a synthetic and geo-mapped population. It runs each agent through a 6-step neural pipeline (Meta's TRIBE v2) that extracts ROI timeseries across fear/salience, reward, deliberation, social, and attention circuits, computes composite scores (arousal, valence, dominance), and derives a per-agent belief-shift value. K2 then generates reasoning traces explaining why each agent moves. A multi-round swarm propagation model shows how support and pushback evolve across the network, visualized on a live geographic map. Finally, we use Tavily and Firecrawl to automatically pull live research dossiers and source excerpts, which grounds each simulation in up-to-date evidence rather than static inputs.

How we built it

Backend: FastAPI orchestrator with SQLite persistence, a vendored 6-step TRIBE neural pipeline, LFCM-style BSV calibration, and K2 batch reasoning.
Frontend: Vite + React + TypeScript with Mapbox, Deck.gl for geo visualization, and Recharts for round history. AI integrations: ElevenLabs for audio transcription, Tavily/Firecrawl for live research dossiers, Modal for optional remote TRIBE batch inference.

Challenges we ran into

Personalizing TRIBE outputs per agent without running a full forward pass for each — we solved this by deep-copying baseline ROI stats and amplitude-scaling from demographic and role features before recomputing composites.
Keeping swarm propagation numerically stable across many rounds and many agents with neighborhood influence factored in.
Getting the frontend map, brain viz, and agent inspection panel to stay in sync with live simulation state.

Accomplishments that we're proud of

A working end-to-end pipeline from raw audio evidence → cortical neural readout → per-agent belief dynamics → geo-propagation map, all in one run.
Per-agent K2 reasoning traces that actually reflect each agent's neural composite state, not just generic LLM opinions.
The TRIBE personalization approach — same stimulus, meaningfully different neural profiles across the population.

What we learned

Mass neural personalization requires precise approximation; doing full passes for each agent is not practical, but amplitude scaling by demographics works well enough to maintain variance.
Swarm behavior is very dependent on neighbor graph structure.

What's next for Cortexia

Longitudinal data sets for the validation of synthetic population calibration.
More complex agents' social networks than just geographic neighborhood effects.
Domain specific training for TRIBE models in health, politics, and finances domains.
Export simulation reports for researchers and policy-makers.