Signa

Signa unifies your blood tests, wearable data and medical records into one AI-powered health companion that cross-analyzes signals no single doctor would catch. 100% Mistral-powered.

Comment

Signa

Your health data should talk to each other so your body doesn't have to scream.

Team

Name LinkedIn Profile Link
Robin Zamora https://www.linkedin.com/in/robin-zamora-4a7302133/
Paul Demarecaux https://www.linkedin.com/in/powl?originalSubdomain=fr

The Problem

70% of blood test results are never discussed in depth with a doctor. Meanwhile, your Apple Watch collects 500 data points per day that nobody analyzes. Despite €400M+ invested in 15+ European preventive health startups, no product combines a health vault, voice AI, and cross-analysis between lab and wearable data not in France, not in Europe.

Your blood panel lives in a PDF in your email. Your wearable data lives in a siloed app. Your doctor sees one, not the other. The signals that matter the ones that predict burnout, chronic inflammation, or metabolic drift only emerge when you cross these sources. Nobody does.

What It Does

Signa is a voice-first health companion that connects the dots between your lab results and your wearable data.

  1. Scan — Take a photo of any French blood panel. Mistral OCR 3 extracts every biomarker in seconds, stores it in your personal health vault.
  2. Cross — Signa crosses your lab biomarkers with your wearable data (heart rate, HRV, sleep, steps via Thryve — 500+ devices supported). Five peer-reviewed cross-signal rules detect patterns invisible to either source alone:
    • HRV + hs-CRP → chronic stress detection (Hartaigh et al., 2016)
    • Ferritin + sleep quality → explained fatigue (PMC4480468)
    • Triglycerides/HDL ratio → insulin resistance screening (AUC 0.849)
    • Resting HR baseline deviation → pre-symptomatic infection detection (Stanford/Mishra, 2020)
    • Steps + glucose stability → positive metabolic signal
  3. Score — A weighted health score + PhenoAge biological age (Levine et al., 2018 — 9 routine biomarkers → biological age and 10-year mortality risk). "You're 38, but your body is 32.5."
  4. Talk — Ask Signa anything about your health by voice. Voxtral STT transcribes, Mistral Small 4 reasons with your full health context injected, Voxtral TTS responds in natural French. The assistant knows your data — it's not a generic chatbot.
  5. Alert — A notification center pushes proactive insights: cross-signal alerts, wearable trend deviations, action reminders ("Your last panel was 6 months ago"), and contextual recommendations aligned with the Alan ecosystem.
  6. Switch — Three real Thryve demo profiles (multi-source student with Oura+Withings+Huawei, active gym user with WHOOP, sedentary tech worker with Apple Watch) demonstrate how Signa adapts to radically different health narratives.

The voice button is context-aware: tap it from the biomarker explorer and the assistant already knows what you're looking at.

Tech Stack

Layer Technology
Framework Next.js 15 + React 19 + TypeScript
Design Tailwind CSS v4 + Framer Motion + Atomic Design System (atoms/molecules/organisms)
LLM Mistral Small 4 (mistral-small-latest) 🚀
OCR Mistral OCR 3 (mistral-ocr-latest) 🚀
STT Voxtral Transcribe (voxtral-mini-transcribe-2507) 🚀
TTS Voxtral TTS (voxtral-mini-tts-2603) 🚀
Wearable data Thryve API 🚀 (500+ devices — Garmin, Withings, Oura, WHOOP, Apple, Samsung, Huawei)
Database Supabase 🚀 (PostgreSQL + Storage)
Hosting Vercel 🚀
Biological age PhenoAge algorithm (Levine & Horvath, 2018)
Testing Playwright — 51 E2E tests
Mobile PWA (installable, standalone)

100% Mistral-native AI stack — 4 Mistral models working together: OCR 3 for document extraction, Small 4 for health reasoning, Voxtral for voice input, Voxtral for voice output.

Special Track

  • [ ] Alan Play: Living Avatars
  • [ ] Alan Play: Mo Studios
  • [ ] Alan Play: Personalized Wrapped
  • [ ] Alan Play: Health App in a Prompt
  • [x] Alan Precision

What We'd Do Next

Signa V2 — from reactive dashboard to proactive health companion:

  • Conversational memory — Every interaction enriches the user's health profile. Signa remembers "you mentioned headaches last month" and correlates with new lab results. Inspired by our work on relational memory systems for AI companions (Kagemi framework).
  • Medication tracking — Upload prescriptions via OCR, get gentle reminders ("Did you take your magnesium?"), and track adherence over time.
  • Post-visit triggers — Alan knows when a member visits a doctor. Signa prompts: "You saw Dr. Martin today — want to upload your results or prescription?"
  • Qualitative data layer — Morning check-ins triggered by wearable patterns ("Your HRV dropped 15% — how are you feeling?"). The user's subjective experience crosses with objective biomarkers.
  • Cross-document intelligence — Graph-based linking between blood panels, prescriptions, consultation reports, and wearable patterns across months and years. Find correlations a single document never reveals.
  • Mon Espace Santé integration — FHIR R4 compatible architecture, ready to plug into France's national health data infrastructure (17M activated profiles, 117M biology reports in 2024).
  • European sovereignty — GDPR-native, data hosted in EU (Supabase West EU Paris), fully compatible with the European Health Data Space (EHDS). The European alternative to what Microsoft is building with Copilot Health in the US.

Architecture

┌──────────────────────────────────────────────────────┐
│              SIGNA — Mobile PWA (430px)               │
│  Home    Explorer    Documents    Alerts    Assistant  │
└───────────────────────┬──────────────────────────────┘
                        │
┌───────────────────────┴──────────────────────────────┐
│                   API ROUTES (Next.js)                 │
│  /api/upload    → Mistral OCR 3 + Small 4 pipeline    │
│  /api/score     → PhenoAge + weighted health score     │
│  /api/chat      → Mistral Small 4 + health context     │
│  /api/stt       → Voxtral Transcribe                   │
│  /api/tts       → Voxtral TTS (fr_marie_neutral)       │
│  /api/protocol  → Cross-signals + LLM hybrid protocol  │
│  /api/wearable  → Multi-source collapse + sparklines   │
│  /api/biomarkers→ Latest per canonical + trends         │
└───────────────────────┬──────────────────────────────┘
                        │
          ┌─────────────┼─────────────┐
          │             │             │
    ┌─────┴─────┐ ┌─────┴─────┐ ┌────┴────┐
    │ Supabase  │ │  Mistral  │ │ Thryve  │
    │ 🚀       │ │  🚀       │ │ 🚀      │
    │ PG + Store│ │  4 models │ │ 500+    │
    │ EU-West   │ │  OCR/LLM/ │ │ devices │
    │           │ │  STT/TTS  │ │         │
    └───────────┘ └───────────┘ └─────────┘

Cross-Signal Science

Our cross-signal engine is based on peer-reviewed research, not heuristics:

Signal Sources Crossed Reference
Chronic stress HRV (wearable) × Cortisol (lab) Hartaigh et al., 2016, Annals of Medicine, N=2,064
Explained fatigue Sleep quality (wearable) × Ferritin (lab) PMC4480468, MDPI 2514-183X
Insulin resistance — TG/HDL ratio, PMC10375356, AUC 0.849
Pre-symptomatic alert Resting HR deviation (wearable) × baseline Stanford/Mishra, 2020, Nature BME, N>5,000
Metabolic positive Steps trend (wearable) × Glucose (lab) —
Biological age 9 routine biomarkers (lab) PhenoAge, Levine & Horvath, 2018, Aging

Quality

  • 51 Playwright E2E tests covering every screen, every API route, every interaction, and design system compliance (no shadows, paper white background, semantic health colors)
  • 3 real Cerballiance blood panels validated through the OCR pipeline (20-26 biomarkers extracted per document)
  • 283 real wearable data rows from 8 Thryve QA profiles, normalized and seeded
  • Deterministic post-processing (21 regex rules) to correct recurring LLM hallucinations on biomarker canonical names

Built With

  • 2-people
  • 4-mistral-models
  • claude
  • framer-motion.-backend:-supabase-(auth
  • mistral
  • next
  • postgres).-wearable-data:-thryve-api.-deployed-on-vercel.-we-used-an-ai-native-development-workflow:-paul-(design/product)-wrote-structured-product-briefs
  • react
  • react-19
  • robin-ran-them-through-orchestria-?-our-autonomous-ai-agent-orchestrator-built-on-claude-code-?-which-executed-the-implementation.-12-hours
  • small-4-for-cross-signal-analysis-and-insight-generation
  • storage
  • supabase
  • tailwind-v4
  • vercel
  • voxtral-mini-for-bidirectional-voice-(stt-+-tts).-frontend:-next.js-15
Share this project:

Updates