CortexFlow is a full-stack cognitive signal analysis platform built for fast screening workflows from text and speech. It combines browser-based voice capture, Gemini-powered transcription, deterministic linguistic feature extraction, and a report interface that maps output to domain-level brain-region signals.

• Accepts typed input or recorded speech
• Transcribes audio through Gemini-compatible APIs
• Scores five analysis domains: lexical, semantic, prosody, syntax, and affective markers
• Streams a structured report with confidence notes and non-diagnostic safety framing
• Visualizes biomarker intensity inside an interactive 3D brain workspace
• Supports authentication and report history through Firebase and Supabase

• Monorepo with clear frontend and backend boundaries
• No proprietary or submission-hostile top-level legal artifacts
• Local demo path available with auth bypass for judging environments
• Deployment-friendly split for Vercel frontend and containerized backend

1. The frontend Next.js app handles UI, authentication, audio upload, report history, and proxy API routes.
2. POST /api/transcribe sends recorded audio to Gemini and extracts transcript plus pause-map timing.
3. POST /api/analyze forwards normalized input to the FastAPI backend.
4. The backend service computes deterministic biomarkers, derives cognitive load, and streams the final report payload back to the UI.

• Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS v4, Recharts, Three.js
• Backend: FastAPI, Pydantic, httpx, Python 3.11+
• AI layer: Gemini API for transcription and report-safe summary generation
• Data layer: Firebase Auth and Supabase for user/session persistence
• Deployment target: Vercel for frontend, Docker-capable backend host for API

.
├── backend/
│   ├── main.py
│   ├── requirements.txt
│   ├── pyproject.toml
│   └── Dockerfile
├── frontend/
│   ├── src/
│   │   ├── app/
│   │   ├── components/
│   │   ├── hooks/
│   │   └── libs/
│   ├── public/
│   ├── supabase/
│   ├── package.json
│   └── Dockerfile
├── .gitignore
└── README.md

• Node.js 20+
• Python 3.11+
• A Gemini API key
• Firebase and Supabase credentials if you want full auth and history support

Create frontend/.env.local from frontend/.env.example.

Required fields for a full run:

GEMINI_API_KEY=
GEMINI_API_BASE=https://generativelanguage.googleapis.com/v1beta
GEMINI_TRANSCRIBE_MODEL=gemini-2.0-flash
BACKEND_URL=http://localhost:8000
NEXT_PUBLIC_BACKEND_WAKE_ENABLED=true
NEXT_PUBLIC_DEV_BYPASS_AUTH=true
NEXT_PUBLIC_FIREBASE_API_KEY=
NEXT_PUBLIC_FIREBASE_AUTH_DOMAIN=
NEXT_PUBLIC_FIREBASE_PROJECT_ID=
NEXT_PUBLIC_FIREBASE_STORAGE_BUCKET=
NEXT_PUBLIC_FIREBASE_MESSAGING_SENDER_ID=
NEXT_PUBLIC_FIREBASE_APP_ID=
FIREBASE_PROJECT_ID=
FIREBASE_CLIENT_EMAIL=
FIREBASE_PRIVATE_KEY=
SUPABASE_URL=
SUPABASE_SERVICE_ROLE_KEY=

Run it:

cd frontend
npm ci
npm run dev

Create backend/.env from backend/.env.example.

GEMINI_API_KEY=
GEMINI_API_BASE=https://generativelanguage.googleapis.com/v1beta
GEMINI_TIMEOUT_SECONDS=40
MODEL_DISCOVERY_TTL_SECONDS=900
GEMINI_REASONING_CANDIDATES=gemini-2.0-flash,gemini-1.5-flash
GEMINI_SAFETY_CANDIDATES=gemini-2.0-flash,gemini-1.5-flash
GEMINI_REASONING_MODEL=gemini-2.0-flash
GEMINI_SAFETY_MODEL=gemini-2.0-flash
MIN_WORDS_REQUIRED=3

Run it:

cd backend
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

1. Start the backend on http://localhost:8000.
2. Start the frontend on http://localhost:3000.
3. Use text input or voice recording.
4. Review streamed cognitive analysis, regional activations, and generated report output.

• Frontend routes
  • POST /api/transcribe
  • POST /api/analyze
  • POST /api/wake-backend
  • GET /api/wake-backend
  • Auth and report routes under frontend/src/app/api
• Backend routes
  • GET /health
  • GET /models/recommended
  • POST /analyze

• Manika Kutiyal
• Aditya Verma

• The repository is structured to be judged as a working hackathon project rather than a private internal codebase.
• Authentication can be bypassed in demo setups with NEXT_PUBLIC_DEV_BYPASS_AUTH=true.
• The backend output is explicitly non-diagnostic and intended for screening/demo workflows.