You are not alone — proven with data.
Built for UNIHACK 2026 · March 13-15, 2026
Echo is a privacy-first mental health app that shows you how many people have felt exactly what you're feeling right now — and what helped them get through it.
You type a thought. Echo anonymises it, finds everyone who felt the same, and tells you: "847 people have felt something like this." You scroll through their experiences. Some of them found a way through — and left a note for you.
No chatbot. No social feed. No clinical questionnaire. Just ambient proof that you're not alone, powered by semantic search and private by design.
You type a thought
|
Qwen3.5-0.8B (self-hosted) strips all identifying info
Raw text is discarded immediately
|
NanoGPT humanises the anonymised text (50-60 words)
Never sees your original words
|
Elasticsearch indexes it with a 384-dim sentiment vector
No account ID, no IP, no raw text stored
|
Vector similarity search finds others who felt the same
|
You see the count + their anonymised experiences
Your raw thought stays only on your device (encrypted)
Core invariant: Raw thought text never persists anywhere except your own device.
- Our server breached? Attacker gets emails + theme categories. Cannot read raw thoughts.
- Elasticsearch breached? Attacker gets anonymised thoughts with zero user linkage.
- Your device breached? Only your data. localStorage is encrypted with AES-GCM.
See docs/PRIVACY_SPEC.md for the full breach analysis.
| Name | Role |
|---|---|
| TBD | TBD |
| Layer | Technology |
|---|---|
| Frontend | Next.js 16 (App Router), TypeScript, Tailwind CSS, Framer Motion |
| Mobile | Capacitor (Android APK via WebView) |
| Backend | FastAPI (Python), SQLAlchemy (async), Pydantic v2 |
| Anonymisation | Qwen3.5-0.8B via Ollama (self-hosted SLM) |
| Humanisation | NanoGPT API (qwen3.5-122b-a10b, OpenAI-compatible) |
| Search | Elasticsearch 8.13 (vector similarity, kNN, aggregations) |
| Embeddings | all-MiniLM-L6-v2 (local, 384-dim) |
| Auth | Email + bcrypt (cost=12), JWT (7-day expiry) |
| Infra | Docker Compose (5 services: PostgreSQL, Ollama, Elasticsearch, FastAPI, Next.js) |
- Docker + Docker Compose v2+
git clone <repo-url>
cd echo-unihack-2026
# 1. Configure environment
cp .env.example .env
# Fill in: NANOGPT_API_KEY, JWT_SECRET
# Optionally: ELASTIC_HOST + ELASTIC_API_KEY for Elastic Cloud
# 2. Start all 5 services
docker compose -f infra/docker-compose.yml up -d
# 3. Wait for healthy status (~60-90s on first run, Ollama pulls ~500MB model)
docker compose -f infra/docker-compose.yml ps
# 4. Seed Elasticsearch with demo data
docker compose -f infra/docker-compose.yml exec backend python seed_elastic.py
# 5. Open the app
# Frontend: http://localhost:3000
# Backend API: http://localhost:8000/docs
# Admin panel: http://localhost:8000/adminFor running without Docker, building the Android APK, or troubleshooting, see docs/START.md.
| # | Feature | Description |
|---|---|---|
| 1 | Thought Submission | Tap the logo, type a thought, full AI pipeline processes it |
| 2 | Response Cards | Scroll through others' humanised thoughts with similarity scores |
| 3 | History Panel | Local history with match counts and resolve button |
| 4 | "What Helped" | Mark a thought resolved, share advice — stored and shown verbatim |
| 5 | Personal Trends | Client-side theme frequency, resolution rate, time-to-resolution |
| 6 | Breathing With Others | Logo animation intensity reflects weekly co-presence count |
| 7 | Future You Letters | Write notes to your future self, resurfaced on matching themes |
| 8 | Guardrails of Care | Crisis helplines shown for risk themes (client-side, zero logging) |
| 9 | Topic Exploration | Tap floating bubbles to browse thoughts by theme without typing |
| 10 | Surrounding Topics | Animated topic bubbles on the home screen perimeter |
| 11 | Saved Anchors | Bookmark "what helped" advice locally for future theme matches |
| 12 | Quiet Wins | Celebrates long gaps between theme mentions |
| 13 | Recurrence Pattern | Flags when the same theme appears 3+ times in 14 days |
| 14 | Resolution Aggregates | Shows per-theme stats on how many people shared advice |
| 15 | Advice-First Toggle | Filter results to only cards with "what helped" attached |
See docs/FEATURE_SPEC.md for full specifications.
# Backend (399 pytest tests — no running services needed)
cd backend && python -m pytest -v
# Frontend type check
cd frontend && npx tsc --noEmit
# Playwright E2E (mocks API, no backend needed)
cd frontend && npx playwright testSee docs/TEST.md for demo mode testing and feature-by-feature instructions.
| Document | Contents |
|---|---|
docs/START.md |
Getting started: Docker, env vars, running without Docker, seeding, troubleshooting |
docs/ARCHITECTURE.md |
System architecture, sequence diagrams, Elastic schema, middleware, Capacitor |
docs/PRIVACY_SPEC.md |
Privacy model, data inventory, breach analysis, compliance notes |
docs/FEATURE_SPEC.md |
Complete specification for all 15 features |
docs/TEST.md |
Testing guide, demo mode behaviour, Playwright test inventory |
docs/PITCH_TEMPLATE.md |
Demo script, submission checklist, third-party API list |
CLAUDE.md |
AI assistant context (privacy rules, conventions, repo structure) |
- NanoGPT API (qwen3.5-122b-a10b) — humanisation + theme classification
- Elasticsearch / Elastic Cloud — vector similarity search + aggregations
- Qwen3.5-0.8B via Ollama — self-hosted anonymisation model
- all-MiniLM-L6-v2 — local sentence embeddings (384-dim)
- AI tools used during development: Claude (claude.ai), Claude Code
- AI Solutions Prize (Quantium) — three-stage AI pipeline: SLM anonymisation, NanoGPT humanisation, Elastic vector semantic search
- Best Use of Elastic Technology — vector similarity, sentiment clustering,
search_afterpagination, real-time aggregate counts - Social Impact Prize — mental health, ambient solidarity, zero clinical gatekeeping, built-in safety guardrails
- Best Design — breathing animation, count reveal, card scroll UX, Future You letters
See LICENSE.