Problem
A fall-response system should still detect and escalate an event when a smartphone, cloud inference, or a continuous internet connection is unavailable.
CASE STUDY / 03
Thesis · Edge ML · Go backend · IoT
Smart Elderly Care & Emergency Response
A co-developed wearable prototype that detects falls on-device, sends emergency alerts over GSM, and reports telemetry to a Go caregiver platform.
01
Overview
The system, not just the screen.
This final-year engineering thesis joins embedded inference, low-connectivity emergency communication, a Go API, PostgreSQL, and a React caregiver dashboard in one end-to-end prototype.
My work
Co-developed the group project across the wearable-to-web path, including the Go backend, device telemetry, authentication, role-aware caregiver flows, and system integration.
02
Architecture
How data moves through the system.
- 01Wearable6-axis IMU · 50 Hz · on-device SVM
- 02A9G GSM / GPRSEmergency SMS and event telemetry
- 03Go + Gin APIDevice auth, JWT, RBAC, clean architecture
- 04PostgreSQL + ReactPersistence and caregiver dashboard
03
Engineering
Decisions and constraints.
Selected decisions
- 01
Run the linear SVM on the nRF52840 so raw IMU inference does not depend on a network round trip.
- 02
Use GSM for direct emergency SMS and GPRS separately for event-level backend telemetry.
- 03
Add a 15-second audible cancellation window before an automatic emergency notification.
- 04
Separate device bearer tokens from human JWT authentication and RBAC.
Challenges
- Fitting filtering, feature extraction, and classification into a constrained wearable runtime.
- Coordinating fall-state transitions, cancellation, SMS delivery, and telemetry under intermittent connectivity.
- Designing one model path that could be evaluated on both SisFall and limited real-world trials.
04
Stack & evidence
What it was built with.
- Go
- Gin
- PostgreSQL
- GORM
- JWT
- React 19
- TypeScript
- C++
- scikit-learn
- Docker
90.69%SisFall F1
0.9984SisFall AUC
95.24%Real-world trial F1
2 false alarms across 67 ADL trialsValidation result
Scope note. Academic prototype, not a certified medical device. GPS was identified as future work and is not implemented. Real-world validation was limited and is not a clinical-performance claim.