Healix

Inspiration

Healthcare in Vietnam and across Southeast Asia suffers from three persistent pain points: patients can't interpret their own lab reports, clinicians drown in documentation, and body composition tracking requires expensive equipment. When Qwen released its multimodal models, we saw a chance to solve all three in one unified platform — using a phone camera and a conversational AI to democratize clinical-grade health insights.

What it does

Healix is an AI health companion with three integrated features:

Labs Analyzer — Upload a photo or PDF of any lab report. Qwen-VL extracts every value, cross-references 50+ medical reference ranges, and classifies each result by severity (normal, borderline, critical) with plain-language explanations in English or Vietnamese.
Clinical Scribe — Record a doctor-patient conversation. Healix transcribes it in real time and uses Qwen-Max to generate a structured SOAP note (Subjective, Objective, Assessment, Plan) that a clinician can review and sign off in seconds.
Body Scan — Point your phone at yourself. A custom Qwen-VL pipeline delivers 22+ body measurements, posture analysis, and a PPG-based heart rate reading — no wearables, no tape measure, no calibration lab.

How we built it

Frontend: React + Vite, mobile-first UI designed screen-by-screen in mockups before a single component was coded.
Backend: FastAPI on Python 3.9, orchestrating requests to Qwen through the Dashscope API.
Labs Analyzer: Qwen-VL parses scanned reports → a reference-range database scores severity → deep-translator handles Vietnamese ↔ English medical terminology.
Clinical Scribe: Google Speech Recognition for transcription → Qwen-Max for SOAP generation. We pivoted to this split after discovering Qwen-Audio wouldn't accept base64 payloads through the OpenAI-compatible endpoint.
Body Scan: We replaced the commercial Bodygram SDK with our own Qwen-VL pipeline that uses a reference object for calibration, then extracts measurements, posture landmarks, and skin-tone PPG signals for heart rate.

Challenges we ran into

Regional endpoint mismatch. We burned hours on a persistent 401 Unauthorized before realizing we were hitting the China Dashscope endpoint from Vietnam. Switching to dashscope-intl.aliyuncs.com (Singapore) fixed it instantly.
Qwen-Audio incompatibility. Base64 audio didn't work through the OpenAI-compatible SDK, so we rearchitected the Clinical Scribe around Google Speech Recognition while keeping Qwen-Max as the clinical reasoning layer.
Body measurements without hardware. Extracting 22+ measurements from a single phone camera needed a reference-object calibration step and a posture-normalization routine we built from scratch.
Dependency hygiene. Missing deep-translator and dashscope in requirements.txt broke collaborator setups — a tiny fix, a useful lesson.

Accomplishments that we're proud of

Shipped three production-quality AI features in a single hackathon sprint.
Built a custom Qwen-VL body-measurement pipeline that competes with commercial SDKs.
Mobile mockups for every screen — not just the happy path.
First-class bilingual medical content (Vietnamese + English).
Zero API keys ever committed to Git.

What we learned

Multimodal LLMs can collapse entire stacks — OCR, NLU, translation, reasoning — into a single well-prompted call.
Regional endpoints matter. A lot.
When a model path fails (Qwen-Audio via OpenAI-compatible), splitting responsibilities across two simpler calls is often faster than fighting the SDK.
Medical UX is unforgiving: every number needs context, every recommendation needs a disclaimer.

What's next for Healix

Clinician dashboard for reviewing, editing, and signing SOAP notes.
Longitudinal tracking — trend lab values and body measurements over time for preventive insights.
FHIR-compatible export so Healix plugs into existing hospital systems.
On-device inference with a quantized Qwen-VL variant for privacy-preserving offline mode.
Expanded reference database into specialty panels (endocrine, oncology, prenatal).