A multimodal brain-response analysis system, built on Meta's TRIBE v2 brain foundation model and Google's Gemma 4. Upload a clip — get a 3D cortical-activation map plus four parallel narrations from four very different readers (an ISU freshman, a WBEZ science reporter, a Northwestern neurologist, and a Google ML scientist).
Picture a movie theatre. Your brain is the audience: 20,484 people in 20,484 assigned seats, each responsible for a specific job — seeing faces, recognising voices, feeling suspense, processing language. The movie is whatever you upload. TRIBE v2, Meta's brain foundation model, is the high-speed sensor system in every seat — twice per second, it predicts how excited each audience member is going to get, three to five seconds before their reaction visibly peaks. Gemma 4 is the panel of four critics in the back booth: after the screening, all four read the same audience-reaction printout and write their own takes — a chatty freshman, a WBEZ reporter, a Northwestern neurologist, and a Google ML scientist. You see all four side-by-side and pick the voice that sounds like your brain.
┌──────────────────────────────────────────────────────────────────┐ │ Browser / phone │ │ ↓ https://big-apple.scylla-betta.ts.net (Tailscale Funnel) │ ├──────────────────────────────────────────────────────────────────┤ │ Big Apple — M4 Max MacBook Pro · 48 GB unified memory │ │ │ │ FastAPI backend (port 8773) — serves HTML + API directly │ │ ├─ TRIBE v2 (PyTorch on MPS, ~6 GB unified) │ │ │ → 20,484-vertex BOLD prediction at 2 Hz │ │ │ │ │ └─ 4× narrate (parallel, in queue) │ │ ↓ │ │ Inference router (port 8766) │ │ ├─→ Local Ollama (Gemma 4 E4B / 26B / 31B on Metal) │ │ └─→ OpenRouter free tier (cloud failover, $0/token) │ ├──────────────────────────────────────────────────────────────────┤ │ Seratonin (Windows · RTX 5090 · Chicago) — STANDBY │ │ ↳ joins router pool when not gaming, runs same code via CUDA │ └──────────────────────────────────────────────────────────────────┘
The whole demo currently runs on a single MacBook Pro (M4 Max,
48 GB unified). Same Python source code as the 5090; the
cortex.device abstraction picks MPS on Apple
Silicon and CUDA on NVIDIA. When Seratonin is back in the
pool, narration jobs round-robin across both nodes; if both fall over, the
router fails to OpenRouter's free Gemma-4-26B endpoint so the demo URL
never returns a 502.
A WebGL/Three.js scene with per-vertex animation, written by Kimi K2.6 via the Nous Portal during the Mercury sprint. Don't read about it — open the live demo and click around.