MedVerse · Team ADAGARD
MedVerse - Sense the unseen.
★ Selected · NIBM Neo Ventures
A full-stack clinical telemetry platform I build with Team ADAGARD - wearable hardware, on-device signal processing, a trained-model layer, a backend, and clinician-facing apps, engineered end to end.
An engineering and research project. MedVerse is not a regulated medical device; the models are trained on public datasets and are not clinically validated.
01 - The motivation
Wards are reactive, not predictive.
Vitals are spot-checked every few hours and continuous monitoring is ICU-only - so between checks the ward is largely blind. Those public-health gaps are what made the engineering problem worth taking on.
- 83%
- of hospital deaths linked to delayed deterioration detection
- 2.5
- CCU beds per 100,000 - versus 9 in Australia
- 9.6%
- mortality even at top tertiary hospitals
- 36%
- national specialist deficit
02 - What I built
A full stack, sensor to dashboard.
I work across the whole stack - wearable firmware and signal processing, the machine-learning layer, the backend services, and the web and mobile apps.
Hardware
A cardiopulmonary vest (v2) with fully fabricated PCBs - KiCad, Gerbers, BOM, pick-and-place - plus a maternal/abdominal belt sensor stack.
Firmware
ESP32-S3 + NimBLE with on-device DSP (filtering, peak detection, HR/HRV/SpO₂) at sub-200 ms sensor-to-alert.
AI / ML
~8 models trained on 40k+ records from public clinical benchmarks, each with held-out validation reports checked into CI.
Backend
FastAPI + TimescaleDB + Redis + Celery + MQTT, with a LangGraph multi-agent layer across 24 endpoints.
Apps
A React clinical dashboard and a Flutter mobile app (patient / clinician / admin), each with a full offline demo mode.
Integrity
A SHA-256 hash-chained audit ledger, and a provenance CI gate that fails the build if an unvalidated model is shown as trustworthy.
03 - The model layer
Models trained on public benchmarks.
Research models, each trained on a public dataset with a held-out metric. These are engineering artifacts - not clinically validated, and not diagnostic.
Sepsis-risk classifier
A calibrated model trained to flag early deterioration patterns.
PhysioNet CinC-2019 · held-out AUROC 0.77
ECG analysis (MI / arrhythmia)
Trained to surface silent-MI and arrhythmia patterns in ECG.
PTB-XL · AUROC 0.86 / 0.78
Fetal CTG classification
A 3-class CTG model paired with a 4-quadrant piezo sensor array.
UCI CTG · AUROC 0.99
Lung-sound classification
An acoustic model over the heart/lung microphone array.
ICBHI 2017 · AUROC 0.74
Cuffless BP estimation
A pulse-transit-time pipeline with per-patient calibration.
Signal-processing model + calibration
ECG-age & risk index
A regression model fused with published rule-based scores (NEWS2).
PTB-XL + peer-reviewed equations
04 - How it's built
Honest by design, and built end to end.
A validated core, honestly labelled. A small set of models trained on real public benchmarks with held-out metrics, surrounded by an explicitly-labelled layer of published clinical equations and signal-processing proxies - so nothing is dressed up as more than it is.
Provenance enforced in CI.A build-time check fails if an unvalidated model is ever surfaced as trustworthy, and the UI flags anything that isn't held-out-validated.
Resilient by default. On-device DSP and local BLE keep the bedside live through network outages; a full offline demo mode means the app never fails to show.
05 - In numbers
The engineering, in numbers.
- ~8
- ML models trained on public clinical benchmarks
- 40k+
- records used for training & held-out validation
- 24
- backend API endpoints
- <200ms
- sensor-to-alert latency, end to end
Selected · NIBM Neo Ventures
Built end to end. Still building.
Happy to walk through the engineering - the hardware, the models, or the platform.