SOMNI AI - Sleep Health Intelligence System

🏆 TreeHacks 2026 Project | Multi-Agent Clinical Evidence Synthesis

🎯 Overview

SOMNI AI transforms consumer wearable sleep data into actionable health intelligence through autonomous multi-agent scientific analysis. The system detects statistical deviations from personal sleep baselines and maps patterns to peer-reviewed longitudinal research associations.

Core Thesis: Sleep deviation is an early biomarker of systemic health drift. SOMNI AI detects trajectory shifts before clinical disease manifests.

Key Features

• 📊 Statistical Sleep Analysis - Z-scores, trend detection, variability indices
• 🤖 Multi-Agent Orchestration - Claude SDK → OpenAI o1 → PubMed → BrightData → Perplexity
• 📝 Dual Report Generation - Patient-friendly (8th grade) + Clinical (evidence-graded)
• 🔬 Scientific Rigor - GRADE methodology, effect sizes, confidence intervals
• 🚀 Production Architecture - Next.js 14 App Router, FastAPI, Edge streaming

🚀 Quick Start

Prerequisites

• Node.js 18+
• Python 3.11+
• API Keys (Anthropic, OpenAI, Perplexity, BrightData)

Installation

# Install Node.js dependencies
npm install

# Install Python dependencies
cd python
pip install -r requirements.txt
cd ..

# Set up environment variables
cp .env.example .env
# Edit .env with your API keys

Running the Application

# Terminal 1: Start Python backend
cd python
python api/main.py

# Terminal 2: Start Next.js frontend (in project root)
npm run dev

Open http://localhost:3000

📁 Project Structure

somni-ai/
├── app/                        # Next.js 14 App Router
│   ├── page.tsx                # Landing page with upload
│   ├── analysis/[id]/page.tsx  # Analysis results
│   └── api/                    # API routes
├── lib/
│   ├── agents/                 # Multi-agent orchestration
│   └── reports/                # Report generators
├── python/
│   ├── sleep_analysis/         # Statistical analysis engine
│   └── api/                    # FastAPI backend
├── components/ui/              # shadcn/ui components
└── public/demo_data/           # Sample datasets

🏆 Prize Narratives & Demonstrations

Greylock — 5+ API Agent with Feedback Reasoning

Criterion: Best hack with an agent that reasons about feedback to dynamically complete complex tasks.

How we meet this:

• 5+ API sequence: Claude (orchestrator) → OpenAI o1 (medical reasoning) → PubMed (literature) → BrightData (guidelines) → Perplexity Sonar (consensus)
• Feedback loop: assess_evidence_quality tool evaluates consistency; returns refine_query, broaden_query, or proceed. When conflicts detected, Claude refines search and re-queries.
• Visible proof: "Multi-API Pipeline" card on analysis page shows all services used. Terminal logs show [Pipeline] sequence and [Feedback] refinements.

Demo points: "Our agent autonomously chooses which API to call and when. When evidence is inconsistent, it refines the query and re-searches before writing reports."

---

OpenAI — Most Creative Use of API

Criterion: Top three teams that use OpenAI API most creatively.

How we meet this:

• o1 as medical reasoning engine: Dedicated system prompt frames task as research associations and risk domain ranking (not diagnosis).
• Structured reasoning: Ask for step-by-step reasoning trace, then structured output (ranked domains, confidence, screening, reasoning).
• reasoning_effort: 'high': Uses o1's extended reasoning for medical analysis.

Demo points: "We use o1 creatively as a clinical reasoning engine with a specialized system prompt. It stays strictly in research associations while providing explicit reasoning traces."

---

Anthropic Human Flourishing

Criterion: Patient report reads empathetically (test with family). Show how AI serves human potential.

How we meet this:

• Empathetic tone: Opening validates effort ("Taking an interest..."), normalizes variation ("Many people see temporary changes").
• Family-tested design: Report designed for reading with family members (see docs/human-flourishing-note.md).
• No alarm: 8th-grade reading level, hopeful framing, clear disclaimers.

Demo points: "We designed the patient report to be read with a parent or partner — supportive, clear, and without fear. AI should make life better by making health intelligence accessible."

---

Anthropic Best Agent (Claude SDK)

Criterion: Autonomous AI application that tackles real problems.

How we meet this:

• Autonomous tool use: Claude agent chooses when to call each of 5 tools based on evidence quality.
• Real problem: Turning messy wearable data into trustworthy, evidence-based health intelligence.
• No hand-coded sequence: Agent dynamically plans pipeline each run.

Demo points: "The orchestrator is a Claude agent with 5 tools. It autonomously decides the pipeline, reasons about evidence, and only then writes reports."

---

OpenEvidence — Clinical Info & Healthcare Track

Criterion (Clinical Info): Convert clinical data into product that improves understanding/decision-making/care delivery. Criterion (Healthcare): Most innovative healthcare product.

How we meet this:

• Clinical data sources: PubMed (literature), BrightData (guidelines), Perplexity (consensus).
• Dual products:
  1. Patient report: Improves understanding (empathetic, actionable, clear next steps)
  2. Clinical report: Supports decision-making (GRADE evidence, screening suggestions, differential diagnosis)
• Innovation: Multi-agent AI + wearable data + evidence synthesis, strictly non-diagnostic.

Demo points: "We convert existing clinical data into two reports: one for patient understanding, one for clinician decision-making. That's healthcare innovation with clear boundaries."

---

TreeHacks Most Impactful

Criterion: Potential to create significant positive change or address pressing societal issue.

How we meet this:

• Early warning system: Sleep deviation as biomarker of systemic health drift before clinical disease.
• Preventive focus: Detect trajectory shifts early so people and providers can act sooner.
• Equity angle: Consumer wearable analysis accessible where sleep labs are scarce.

Impact statement: "SOMNI AI turns wearable data into evidence-based insights so people can spot patterns before they become disease — reducing healthcare burden and supporting preventive, equitable access to understanding."

Demo points: "We're not diagnosing — we're surfacing research-backed patterns early. That's preventive, person-centered, and reduces pressure on the system."

---

Vercel — Best Use of Vercel

Criterion: Best leverages Vercel to build, deploy, and scale production-ready app.

How we meet this:

• Single deployment: TypeScript sleep analysis (no Python backend required) — everything on Vercel.
• Next.js 14 App Router: Server components, API routes, serverless functions.
• Production-ready: Full deployment with caching and edge capabilities.

Demo points: "The whole app runs on Vercel — frontend and APIs. We ported sleep analysis to TypeScript so there's no separate backend. One deploy, production-ready."

---

BrightData — Best AI-Powered Web Hack

Criterion: Use BrightData in an innovative way.

How we meet this:

• Innovative use: Pull public health guidelines and disparity data (CDC/AHA-style) into agent pipeline.
• Value-add: Grounds recommendations in current guidelines; surfaces population-level context (disparities).
• Real usage: BRIGHTDATA_API_KEY set in production; "BrightData: real" shown in UI.

Demo points: "BrightData feeds our agent with guideline and disparity datasets so recommendations align with public health guidance."

---

Perplexity — Best Use of Sonar API

Criterion: Build something extraordinary with Sonar; innovative use, technical excellence, real-world value.

How we meet this:

• Consensus + controversy: Agent asks Sonar for both consensus AND controversy in literature.
• Real-world value: Calibrates strength of claims in reports based on where evidence agrees/disagrees.
• Evidence synthesis: Sonar citations integrated into research foundation.

Demo points: "We use Perplexity Sonar as our consensus engine: we ask for consensus and controversy so reports reflect where the evidence agrees and where it doesn't."

📚 Using Demo Data

Test the system with the included sample:

# Use the 30-day Fitbit sample
public/demo_data/fitbit_30days.csv

Upload this file through the web interface to see the full analysis pipeline.

⚠️ Important Disclaimers

SOMNI AI is NOT a diagnostic tool. It:

• Does NOT diagnose medical conditions
• Does NOT provide medical advice
• Uses consumer wearable estimates (not clinical PSG)
• Shows research associations, not causation

Always consult healthcare providers for medical decisions.

🧪 Testing

# Python tests
cd python
pytest tests/ -v

# Type checking
npm run type-check

# Build verification
npm run build

📧 Contact

Built for TreeHacks 2026

---

MIT License - See LICENSE file