An AI-powered web app that helps patients organize their symptoms before a doctor's visit. The core design constraint: zero data persistence — no database, no user accounts, no logs containing health data.
Patients arrive at consultations anxious and forget key details. Doctors have limited time. The gap between what the patient knows and what the doctor hears costs both sides.
PreConsult bridges that gap with a guided AI interview that generates a structured clinical summary — then destroys all data when the session ends.
- Patient arrives at a welcome screen explaining the process with a clear privacy reassurance
- Patient fills out demographics (age, sex) and their chief complaint with duration
- The AI generates targeted follow-up questions using clinical frameworks (OPQRST + SAMPLE)
- Answers are compiled into a structured summary the patient can download as a PDF
- Session data is destroyed — nothing is retained on the server
┌──────────────────────────────────────────┐
│ PreConsult App │
│ (Unified Cloud Run Service) │
├────────────────────┬─────────────────────┤
│ Reflex UI │ FastAPI Backend │
│ (React/Next.js) │ (Interview/PDF) │
└─────────┬──────────┴──────────┬──────────┘
│ │
▼ ▼
┌───────────────────┐ ┌───────────────────┐
│ Redis Session │ │ Vertex AI │
│ (30-min Context) │ │ (Clinical LLM) │
└───────────────────┘ └───────────────────┘
No health data is ever written to disk. Every design decision flows from this constraint:
- No database — session state lives in Redis with a 30-minute TTL.
- No user accounts — complete anonymity, no registration required.
- In-memory PDF generation — reports are built in RAM and streamed directly to the browser.
- No data in transit to disk — even if Redis fails, rate limiting falls back to in-memory counters (no clinical data).
- No model training — API calls use contracts that exclude session data from training.
| Layer | Technology |
|---|---|
| Core | Reflex 0.9.1 (Unified Frontend & API Host) |
| Backend | FastAPI (Integrated into Reflex backend) |
| Session | Redis (Hashes, ephemeral, 30-min TTL, RedisUnavailable → 503) |
| AI | Vertex AI Gemini 2.5 Flash Lite (via LangChain) |
| ReportLab (in-memory, deterministic, no LLM call) | |
| UI/UX | Glassmorphism, mobile-first, 48px touch targets, prefers-reduced-motion, EN/PT i18n |
| Monitoring | Sentry (error tracking) + GET /health endpoint |
| Deployment | GCP Cloud Run (1.0 CPU, 1Gi RAM, 0-5 instances) |
| CI/CD | GitHub Actions (tests + WIF auth + Cloud Run deploy) |
Early versions used ChromaDB and a VM-based RAG pipeline. We replaced it with specialized clinical prompt engineering using established frameworks (OPQRST for symptom assessment, SAMPLE for medical history).
Foundation models already contain the necessary medical knowledge. Structured prompting yields cleaner, faster results at a fraction of the infrastructure cost — eliminating ~$100/month in cloud spend.
The project uses uv for dependency management.
- Python 3.11
- Redis (via
docker composeor local install) - GCP service account with Vertex AI access
.envfile in the project root
docker compose up -d redis
uv sync
uv run reflex runCreate a .env file:
PRECONSULT_API_KEY=your_key
GOOGLE_APPLICATION_CREDENTIALS=path/to/credentials.json
GOOGLE_CLOUD_PROJECT=your_project_id
REDIS_URL=redis://localhost:6379/0
SENTRY_DSN= # optional
GTAG_ID= # optionaluv sync --extra test
uv run python -m pytest tests/ -vThe suite has 50 tests covering:
- Agent service (LLM chains, language, emergency detection)
- API integration (session init, streaming, PDF generation, analytics)
- PDF generation (i18n, pagination, empty states)
- Rate limiting (Redis + in-memory fallback, concurrency)
- Security (API key enforcement, error sanitization, BUILD_MODE)
- Session service (CRUD, Redis recovery after failure)
- i18n (fallback, key consistency,
get_localized_value)
| Method | Path | Description |
|---|---|---|
GET |
/ |
Root — welcome message |
GET |
/health |
Liveness probe (returns {"status": "healthy", "redis": "ok"}) |
POST |
/api/session/init |
Create session with full form data (rate limited: 2/min) |
POST |
/api/interview-questions-stream |
SSE stream of LLM-generated clinical questions |
POST |
/api/generate-pdf |
Deterministic PDF from form + Q&A (no LLM call) |
POST |
/api/analytics/event |
Log analytics event |
GET |
/api/analytics/stats |
7-day funnel stats (token-gated) |
The app deploys as a single consolidated container via GitHub Actions CI/CD.
gcloud run deploy preconsult \
--source . \
--project=securemed-chat-494521 \
--region=southamerica-east1 \
--memory=1Gi --cpu=1 \
--min-instances=0 --max-instances=5 \
--concurrency=80 \
--set-secrets=PRECONSULT_API_KEY=PRECONSULT_API_KEY:latest,REDIS_URL=REDIS_URL:latestThe app is served at pre-consult.org via a Cloudflare Worker that proxies requests to the Cloud Run endpoint, rewriting the Host header to match GCP's requirements.
preconsult_chat/
├── src/preconsult/ # Backend package
│ ├── api/endpoints.py # FastAPI routes
│ ├── core/config.py # Environment config
│ ├── core/errors.py # Exception handlers
│ ├── core/llm.py # Vertex AI singleton
│ ├── main.py # FastAPI app + Sentry + health
│ └── services/
│ ├── agent_service.py # LangChain prompts + streaming
│ ├── pdf_service.py # ReportLab PDF generation
│ └── session_service.py# Redis + in-memory rate limiting
├── reflex_app/preconsult/ # Reflex frontend
│ ├── preconsult.py # UI components + app setup (7 step wizard)
│ ├── state.py # Reflex state + API calls
│ ├── analytics.py # Analytics via HTTP
│ └── i18n.py # EN/PT translations (190+ keys)
├── tests/ # 50+ tests
├── Dockerfile # Multi-stage build
├── docker-compose.yml # Local Redis
└── .github/workflows/ci-cd.yml
| Phase | Focus | Commits |
|---|---|---|
| 0 | Security: race condition, API_BASE_URL, BUILD_MODE, Sentry | 1 |
| 1 | Resilience: Redis fallback, error handlers, sanitized errors, analytics HTTP | 2 |
| 1.5 | Test coverage: mock cleanup, 7 new tests | 1 |
| 2 | Maintenance: gender-neutral, PDF i18n, Dependabot | 1 |
| 3 | Coverage: concurrency tests, i18n edge cases | 1 |
| 4 | Infra: healthcheck, MockWebSocket extraction | 1 |
| 5 | UI/UX: mobile-first layout, inline errors via callout, streaming timeout | 2 |
| 6 | UI/UX: landing split, privacy trust, accessibility for fragile users | 1 |
| 7 | UI/UX: chief complaint reorder, "none" toggle, emergency dialog, skeleton | 1 |
| 8 | UI/UX: prefers-reduced-motion, conditional lang selector, error callout | 1 |
| Total | 50+ tests |
The UI was redesigned with three core principles:
- Mobile-first: All touch targets are 48px minimum, spacing is generous (16px+), and grids collapse to 2 columns on mobile for easier tapping
- Privacy as trust: A privacy badge appears on the landing page and each step, reassuring users that no data is stored
- Fragile-user friendly: Font sizes start at 16px for body text, animations respect
prefers-reduced-motion, skeletons replace spinners during loading, and error callouts use high-contrast solid styling
Disclaimer: PreConsult is an organizational tool, not a medical device. It does not provide diagnoses or medical advice. Always consult a qualified healthcare provider.