From you, to you.
A private, AI-powered daily practice that helps employees understand their own patterns.
Local-first. Privacy-first. Containerized with Podman. Zero cloud required.
Pods, Prompts & Prototypes 2026 · Challenge 1: "My First Local AI" (Beginner Tier)
Quick Start · X-Ray Mode · MCP Tools · Employer Dashboard · Ethics
podman run -p 8000:8000 docker.io/odominguez7/yu-shield:latestOpen http://localhost:8000. Click "Load Demo Data" to populate a 530-person organization (9 departments, 14 days of check-in history).
All AI runs locally via IBM Granite 3.3 through RamaLama. No data leaves your machine.
Employee burnout costs U.S. employers $125-190 billion annually. Current solutions (EAPs, wellness apps) are reactive. They wait for employees to self-refer after a crisis. Only 3-5% of employees ever use them.
YU Shield uses behavioral intelligence to detect burnout signals before they become crises. Through a 30-second daily check-in, YU builds a personal behavioral baseline. When patterns shift (not a bad day, but sustained deviation over 3+ days), the AI delivers personalized, CBT-informed micro-interventions grounded in the employee's actual data.
Key insight: Taking care of yourself is not selfish. It is one of the best ways to take care of the people who depend on you.
┌─────────────────────────────────────────────────────────┐
│ Browser (React SPA) │
│ Landing · Check-In · Dashboard · MCP Playground │
│ X-Ray Mode · Wellness Lab · Activity Booking │
└───────────────────────┬─────────────────────────────────┘
│ REST API
┌───────────────────────▼─────────────────────────────────┐
│ Podman Container (:8000) │
│ │
│ FastAPI Backend │
│ ├── 12 REST endpoints │
│ ├── SQLite database (users, check-ins, departments) │
│ ├── Drift detection engine (baseline + 3-day rolling) │
│ ├── FastMCP server (5 wellness tools) │
│ └── Dual AI provider routing │
│ │
└────────┬────────────────────────────────┬───────────────┘
│ │
┌────────▼────────┐ ┌──────────▼──────────┐
│ LOCAL (default) │ │ CLOUD (X-Ray only) │
│ RamaLama │ │ Cloud AI (compare) │
│ Granite 3.3 8B │ │ (optional) │
│ Data stays here │ │ For comparison only │
└─────────────────┘ └─────────────────────┘
Same check-in. Two models. One runs locally on Granite 3.3 via RamaLama. The other hits a cloud API. The result? You won't miss the cloud. Local AI delivers enterprise-quality coaching with zero data leaving your infrastructure. That's the whole point of X-Ray: seeing is believing.
┌─────────────────────────┐ ┌─────────────────────────┐
│ LOCAL Granite 3.3 │ │ CLOUD Cloud AI │
│ Data never left device │ │ Sent to cloud API │
│ │ │ │
│ "Your mood dropped │ │ "I notice your mood │
│ from 4.2 to 2.6..." │ │ has shifted from..." │
│ │ │ │
│ 2.1s │ │ 1.8s │
└─────────────────────────┘ └─────────────────────────┘
Personal Baseline (7+ days) → Rolling 3-day Average → Drift Alert
mood: 4.2 mood: 2.3 -1.9 below baseline
energy: 4.1 energy: 2.0 -2.1 below baseline
sleep: 4.3 sleep: 2.0 -2.3 below baseline
Every AI response cites specific data points and includes a confidence indicator:
- Limited data (< 7 check-ins): orange indicator
- Building confidence (7-13 check-ins): amber indicator
- Strong baseline (14+ check-ins): green indicator
5 tools exposed via Model Context Protocol. Any MCP-compatible AI assistant can interact with YU Shield:
| Tool | Description |
|---|---|
check_my_wellness |
Employee self-service: check your own baseline, trends, drift |
get_team_wellness_summary |
Anonymous team aggregates (no individual data) |
book_wellness_activity |
Book from catalog of 17 activities |
get_wellness_recommendations |
Score-based activity suggestions |
submit_checkin |
Submit check-in and get AI coaching response |
Live interactive MCP Playground at /mcp for testing tools in real-time.
Enterprise-grade organizational wellness view for Meridian AI (530-person demo company):
- Department-level breakdown: 9 departments with headcount, avg scores, trend arrows, drift alerts, participation rates
- 14-day trend charts: Area charts for mood, energy, sleep
- Environment alerts: Drift detection surfaced as organizational signals, not individual blame
- Actionable insights: Environmental recommendations (reduce meeting load, review workload distribution)
- Zero individual data. Anonymous aggregates only. Individual access is architecturally impossible.
16 science-backed protocols + 5 team challenges with study citations:
- Cold Plunge Duo (250% dopamine increase. Sramek et al., 2000)
- Pre-Meeting Breathwork (62% cortisol reduction. Ma et al., 2017)
- Walking 1:1 Challenge (60% more creative ideas. Stanford, 2014)
17 activities across 4 categories (Calm, Energize, Focus, Recovery) with named local providers, durations, locations, and intensity levels.
1. Open app → Landing page: "From you, to you."
2. Enter first name → Registration (no email, no password)
3. Daily check-in → 30 seconds: mood + energy + sleep + optional note
4. AI responds → Personalized coaching citing your data
5. Approve response → "This resonates" / "Not helpful" (human-in-the-loop)
6. Book activity → Browse and book wellness activities
7. Build baseline → 7+ days creates personal behavioral baseline
8. Drift detection → AI alerts when patterns shift, suggests interventions
9. Employer sees → Only anonymous team aggregates. Never individual data.
podman run -p 8000:8000 docker.io/odominguez7/yu-shield:latest# Clone
git clone https://github.com/odominguez7/podman-hackathon.git
cd podman-hackathon
# Build and run with Podman
podman build -t yu-shield .
podman run -p 8000:8000 \
-e RAMALAMA_URL=http://host.containers.internal:51564 \
yu-shield# Terminal 1: Start local AI model
ramalama serve granite-3.3-8b-instruct
# Terminal 2: Backend
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
RAMALAMA_URL=http://localhost:51564 uvicorn app.main:app --reload --port 8000
# Terminal 3: Frontend
npm install && npm run devClick "Load Demo Data" on the Dashboard, or:
curl -X POST http://localhost:8000/api/seed-demoCreates 38 users across 9 departments with 14 days of realistic check-in patterns (burnout arcs, recovery trends, stable baselines).
| Endpoint | Method | Description |
|---|---|---|
/api/users |
POST | Register user |
/api/checkin |
POST | Submit check-in + get AI response |
/api/checkin/compare |
POST | X-Ray Mode: both providers with timing |
/api/insights/{user_id} |
GET | AI wellness analysis |
/api/history/{user_id} |
GET | Check-in history + baseline |
/api/dashboard |
GET | Anonymous team aggregates |
/api/dashboard/team |
GET | Anonymized per-member status |
/api/dashboard/departments |
GET | Department-level wellness stats |
/api/mcp/recommendations |
POST | Activity recommendations (MCP) |
/api/mcp/book |
POST | Book wellness activity (MCP) |
/api/seed-demo |
POST | Load enterprise demo data |
podman-hackathon/
├── app/ # Python backend
│ ├── main.py # FastAPI. 12 endpoints + SPA serving
│ ├── database.py # SQLite. Users, check-ins, baselines, drift, departments
│ ├── shield.py # AI engine. Dual provider (local + cloud)
│ ├── mcp_server.py # FastMCP. 5 wellness tools
│ └── seed_demo.py # Enterprise demo data (38 users, 9 departments)
├── src/ # React frontend
│ ├── pages/
│ │ ├── Landing.tsx # Hero page: "From you, to you."
│ │ ├── Chat.tsx # 4-tab UI: Check-in, Insights, Lab, Book
│ │ ├── Dashboard.tsx # Enterprise dashboard (Meridian AI, 530 employees)
│ │ └── MCPPlayground.tsx # Live MCP tool testing
│ ├── components/
│ │ ├── WellnessHub.tsx # Analytics + Wellness Lab protocols
│ │ ├── BookingInline.tsx # Activity catalog with booking
│ │ └── BookingModal.tsx # Modal booking (17 activities, 4 categories)
│ └── lib/
│ └── api.ts # API client
├── Dockerfile # Multi-stage: Node 20 build, Python 3.12 runtime
├── docker-compose.yml # Podman Compose with volume persistence
├── requirements.txt # FastAPI, FastMCP, httpx, Pydantic
├── .env.example # Environment variable template
└── SUBMISSION.md # 1-page hackathon submission
Built for the employee. Not the employer.
Individual data is architecturally inaccessible to employers. Not policy-blocked. Structurally nonexistent. Employees own their data. Employers see only anonymous team aggregates (minimum 5 people per group). No names. No individual scores. No check-in notes. Ever.
With RamaLama, all AI runs on-device. Data never leaves the machine.
YU Shield was designed with the hackathon's ethics criteria in mind:
| Ethical Principle | Implementation |
|---|---|
| Human approval step | "This resonates" / "Not helpful" feedback after every AI response |
| Evaluation metrics | Confidence indicator (Limited / Building / Strong) based on data points |
| Personalization with constraints | Baseline requires 7+ check-ins before personalized recommendations activate |
| Grounded outputs | Every AI claim cites specific data points from the user's history |
| Uncertainty handling | AI never makes clinical claims. Frames everything as wellness support |
| Professional referral | When scores are critical, AI recommends EAP or wellness professional |
| Data ownership | Users control their own data. Export and deletion supported |
| Architectural privacy | Employer access to individual data is structurally impossible |
| Layer | Technology |
|---|---|
| Container | Podman + multi-stage Dockerfile |
| Local AI | RamaLama + IBM Granite 3.3 8B Instruct (open source, Apache 2.0) |
| Cloud AI | Cloud AI provider (X-Ray mode comparison only) |
| MCP | FastMCP (5 tools, live playground) |
| Backend | FastAPI + SQLite + Python 3.12 |
| Frontend | React 18 + Vite + TypeScript + Tailwind CSS + shadcn/ui |
| Charts | Recharts |
| Registry | docker.io/odominguez7/yu-shield:latest |
| Criteria | How YU Shield Addresses It |
|---|---|
| User Flow & UI/UX | 3-step check-in (select mood, set energy/sleep, submit). Clear entry point. Keyboard shortcuts. Responsive design. |
| Engineering Quality | Multi-stage Dockerfile, dual AI providers with graceful fallback, SQLite with indexes, MCP server, department aggregation engine |
| Grounding & Verifiability | Every AI response cites data points. Confidence indicators. Baseline/drift numbers visible. Sources shown. |
| Creativity & Innovation | X-Ray Mode (local vs cloud arena). Department-level drift detection. MCP Playground. Human approval loop. |
| Use Case & Real-World Fit | Employee burnout is a $190B problem. YU targets it with 30-second daily check-ins. Privacy-first for regulated industries. |
I lost 80 lbs, ran the Boston Marathon, and finished Ironman 70.3. Not because I had perfect conditions, but because I learned to protect my energy before everything else.
YU Shield exists because the people who depend on you need you at your best.
Pods, Prompts & Prototypes 2026
The Open Accelerator, Fort Point, Boston
Omar Dominguez · MIT Sloan Fellow MBA · [email protected]
