A B2B IoT + Edge AI SaaS platform that monitors physiological stress signals of blue-collar workers in real time — predicting accidents before they happen. Built on PPG + EDA + Thermal + TinyML fusion. Targeted at India's construction, manufacturing, and mining MSME sectors.
After scoring every technology in the PDF across VC traction, research momentum, Indian industrial applicability, Edge AI depth, and build feasibility — the highest-scoring stack isn't a single sensor. It's a physiological signal fusion built from four complementary technologies that individually cost under ₹1,300 total but together create a sensing capability no existing Indian product has deployed.
Heart rate, SpO₂, HRV. The HRV signal alone is a validated predictor of fatigue, heat stress, and cardiovascular collapse risk. This is the anchor signal.
Electrodermal activity from sweat gland activation. Captures sympathetic nervous system stress response that precedes physical collapse by 8–12 minutes.
Contactless core temperature proxy via wrist skin temperature. Combined with ambient heat data, constructs a real-time heat strain index per worker.
A quantized neural network running on ESP32 that classifies the fused signal into Low / Medium / High / Critical risk states in under 50ms — offline, private, instant.
EMG gives localized muscle fatigue but requires gel electrodes repositioned by skill — impractical at construction sites. EEG is prohibitively complex for continuous wear. Bioimpedance needs careful calibration per individual. PPG + EDA is wrist-based, passive, continuous, and validated across thousands of published studies. The signal fusion is what competitors haven't done.
| Technology | VC Signal (2025) | India Fit | Build Feasibility | Edge AI Depth | Score |
|---|---|---|---|---|---|
| PPG + EDA + Thermal Fusion ✓ | 🔥 Very High | 🔥 Perfect | 🔥 Easy | 🔥 Deep | 9.4 / 10 |
| WiFi CSI (Presence/Breathing) | High | Medium | Medium | High | 7.8 / 10 |
| EMG (Muscle Fatigue) | High | Low | Hard | High | 6.9 / 10 |
| Seismic / Vibration | Medium | High | Easy | Medium | 6.5 / 10 |
| Doppler Radar | Medium | Medium | Medium | Medium | 6.2 / 10 |
Blue-collar workers in construction (56M workers), manufacturing MSMEs (110M), and mining (2.5M in Odisha, Jharkhand, Chhattisgarh alone). These are not isolated incidents — India averages 3 deaths and 11 injuries in registered factories every single day, and this covers only 10% of the actual workforce. The informal sector is completely invisible.
A construction worker in 42°C heat with 80% humidity crosses a critical Heat Strain Index threshold roughly 2.3 hours into a shift. Nobody knows. A supervisor walking by cannot see it. The worker's body suppresses distress signals cognitively — until syncope.
Existing industrial safety solutions in India are either:
Not a single affordable wearable in the Indian market does real-time physiological risk scoring. The closest products — Garmin, Apple Watch — are ₹30,000+ and designed for recreation, not industrial fatigue classification.
Construction workers exposed to elevated temperatures experience significant productivity loss — confirmed across 2,387 workers in 2024 meta-analysis (BMC Public Health)
Deaths per 100,000 workers per year in India. Fatality rate 4× higher than the EU industrial average, yet occupational health monitoring market penetration is near zero.
Of India's workforce is in the informal sector — completely outside the safety reporting system. No data, no alerts, no accountability, no product serving them.
No product in India continuously monitors a worker's physiological state (heart rate variability + skin conductance + skin temperature) and generates a real-time predictive risk score visible to both the worker and supervisor — before an incident occurs. The gap isn't technology. The gap is a deployment-ready, affordable, rugged, connectivity-independent system designed for Indian industrial constraints: dust, heat, illiterate end-users, spotty 4G, ₹500/month labour cost context.
VitalEdge is a three-component system: a wrist-worn IoT device, an ESP32-based site gateway, and a cloud SaaS platform. The device continuously reads PPG, EDA, and skin temperature. An on-device TinyML model classifies risk state every 10 seconds. A risk score (0–100) is transmitted via ESP-NOW to the gateway, which aggregates all workers and pushes to cloud. Supervisors see a live floor map. HR managers see weekly physiological compliance reports. Safety officers get predictive incident alerts with 8–15 minute lead time.
| Component | Purpose | Cost |
|---|---|---|
| ESP32-C3 Mini | MCU + BLE + WiFi + ESP-NOW | ₹180 |
| MAX30102 | PPG → HR + SpO₂ + HRV | ₹150 |
| Grove GSR | EDA stress index | ₹300 |
| MLX90614 | Contactless skin temp | ₹300 |
| LiPo 500mAh | 8–10 hr battery life | ₹200 |
| TP4056 module | USB-C charging | ₹30 |
| DRV2605L | Haptic feedback motor | ₹120 |
| 3D printed / silicone band | Rugged wrist enclosure | ₹80 |
| Total | ₹1,360 |
| Component | Purpose | Cost |
|---|---|---|
| ESP32-WROOM-32 | ESP-NOW hub, up to 20 devices | ₹250 |
| SIM7600E 4G module | Cloud uplink via cellular | ₹800 |
| SD Card module | Offline data buffering | ₹80 |
| 12V industrial enclosure | IP65 rated site box | ₹400 |
| DC-DC Buck converter | Power from site supply | ₹60 |
| Total | ₹1,590 |
Edge AI on the device makes instant (10-second) decisions about current risk. Cloud AI learns population-level patterns over weeks and refines per-worker baselines. Together, they become more accurate the longer a customer uses the platform — creating a compounding intelligence advantage that new entrants cannot replicate quickly.
Model: Quantized neural network, INT8, ~95KB. Built in Edge Impulse. Trained on open-access physiological datasets (WESAD, DEAP, SWELL) + synthetic heat-stress augmentation.
Input features (per 10-sec window):
Output: Risk class {0: Safe, 1: Caution, 2: Elevated, 3: Critical} + confidence score.
Inference latency: <50ms. No internet needed. Works in dead zones, tunnels, basements.
Per-worker baseline profiling: First 5 shifts establish individual physiological baseline. Alerts adapt to individual variation — a worker who naturally has high EDA is scored relative to their own norm, not population average.
Shift-level fatigue curve modeling: XGBoost model trained on time-series of risk scores per shift. Predicts when a given worker will breach risk threshold based on current shift trajectory — gives supervisor a 8–15 minute predictive window.
Anomaly detection: Isolation Forest flags physiological patterns inconsistent with work context — e.g., elevated HR + low EDA at rest could indicate cardiac event rather than heat stress. Separate alert pathway.
Population-level insights: Aggregate shift analytics identify which worktimes, temperatures, task types, and site locations produce highest risk — actionable for safety policy, not just monitoring.
Use WESAD (Wearable Stress and Affect Detection) dataset from TU Munich — 15 subjects, PPG + EDA + temperature, stress/non-stress labels. Augment with synthetic heat-stress scenarios using temperature ramp simulation. Target: 85%+ accuracy on 3-class classification (safe/caution/critical) before first pilot. Edge Impulse provides end-to-end pipeline: data ingestion → feature engineering → model training → INT8 quantization → ESP32 deployment library.
The device is the entry point. The SaaS platform is the business. Every additional day of data makes the predictions more accurate. Every additional site makes the population model richer. The SaaS is not a monitoring dashboard — it is a Workplace Physiological Risk Intelligence platform with three distinct user roles.
Floor map of all workers color-coded by real-time risk score. Instant push notification (app + WhatsApp) when any worker enters Elevated/Critical state. Single-tap emergency response log. Works on a ₹8,000 Android phone.
Shift trajectory forecasts. "Worker #14 is predicted to breach Critical threshold in ~12 minutes based on current pattern." Incident prediction history with accuracy scoring. DGFASLI-format compliance export.
Weekly / monthly physiological compliance reports. Risk distribution across departments, shifts, seasons. Comparison of risk profiles pre/post safety interventions. Cost-of-incident estimation (insurance integration ready).
Native mobile apps fail in Indian industrial contexts — supervisors don't check apps, don't install updates, forget passwords. VitalEdge alerts via WhatsApp Business API. A Critical alert sends: "⚠️ Worker Ramesh (Zone B) — Critical Heat Stress Risk. Predicted incident in ~10 min. [Tap to acknowledge]". This single feature removes the biggest deployment friction in India and has no equivalent in any competing product.
| Layer | Technology | Reason |
|---|---|---|
| Backend API | FastAPI (Python) | Native ML integration, async time-series ingestion |
| Database | TimescaleDB (PostgreSQL extension) | Purpose-built for time-series sensor data, SQL compatibility |
| ML Serving | Python + Scikit-learn + XGBoost | Your core stack; familiar, deployable |
| Frontend | React + Recharts | Rapid SaaS dashboard development |
| Cloud | AWS IoT Core + Lambda / GCP IoT | Scales from 10 to 10,000 devices without re-architecture |
| MQTT Broker | AWS IoT Core (managed) | Industry standard IoT message protocol, free tier covers prototype |
| Alerts | WhatsApp Business API (Meta) | 90%+ engagement vs. push notifications in Indian blue-collar context |
| Edge ML | Edge Impulse + TensorFlow Lite Micro | Free, ESP32-native, end-to-end pipeline |
The revenue model separates hardware (one-time or leased) from software (annual contract). The SaaS contract is the recurring, high-margin component. Hardware provides entry-point revenue and creates switching costs. Pricing is anchored to the cost of a single prevented incident — a factory fatality costs a company ₹8–25 lakhs in compensation, regulatory fines, and downtime. The SaaS subscription is ~1% of that number per worker per year.
Hardware: 100 × ₹3,999 = ₹3.99L one-time. SaaS: 100 × ₹799 × 12 = ₹9.59L/year. Year 1 contract value: ₹13.58L per client. Cost of goods (hardware + cloud): ~₹2.8L. Gross margin on SaaS after Year 1: ~78%. Target: 15 enterprise clients by Month 18 = ₹1.44Cr ARR.
Every shift generates labeled physiological data from Indian workers in real conditions. By Month 18, this dataset doesn't exist anywhere else. It enables model performance no competitor can match without years of deployment.
Each worker's 6-month physiological baseline is stored in the platform. Switching vendors means losing all baseline profiles — predictive accuracy drops to Day 1 level. This creates real switching cost.
Once WhatsApp alerts, compliance reports, and ERP API feeds are integrated into daily safety workflows, the product is structurally embedded. Removal requires rebuilding the entire safety protocol.
DGFASLI compliance export positions VitalEdge as the evidence layer for regulatory audits. When occupational health compliance becomes mandatory — and it will — VitalEdge clients are already compliant.
| Company / Solution | Real-time Physiology? | Predictive Risk? | India-deployable? | Under ₹2,000/worker? | Heat Stress Focus? |
|---|---|---|---|---|---|
| VitalEdge (proposed) | ✓ Yes | ✓ Yes | ✓ Yes | ✓ Yes | ✓ Core feature |
| Jarsh Safety (India) | ✗ No | ✗ No | ✓ Yes | ⚠ Partial | ✗ No |
| SlateSafety (US) | ✓ Yes | ⚠ Basic | ✗ Too expensive | ✗ No (₹15K+) | ✓ Yes |
| Generic GPS trackers | ✗ No | ✗ No | ✓ Yes | ✓ Yes | ✗ No |
| Smart helmets (various) | ✗ No | ✗ No | ⚠ Partial | ✗ No (₹5K+) | ✗ No |
| Apple Watch / Garmin | ✓ Yes | ✗ No | ✗ No (₹25K+) | ✗ No | ✗ No |
Solder MAX30102 + Grove GSR + MLX90614 to ESP32 breadboard. Validate all three signals cleanly. Record your own physiological response during exercise, rest, and stress induction. Upload to Edge Impulse. Build first classification model. Target: can you reliably distinguish "resting" vs. "exercising" physiologically? This is your Day 0 proof-of-concept.
Download WESAD dataset. Retrain on combined WESAD + your own data. Target 80%+ accuracy on 3-class risk. Quantize to INT8. Deploy on ESP32. Measure inference time. Implement ESP-NOW transmission of risk score packet. Build the gateway receiver. Whole pipeline must work: wrist → ESP-NOW → gateway → serial print.
Stand up FastAPI + TimescaleDB on AWS Free Tier. Build MQTT ingestion from gateway. Implement per-worker risk score time-series storage. Build shift trajectory model (XGBoost). Implement WhatsApp Business API alert for Critical state. Create basic supervisor dashboard in React. Test end-to-end with 3 wrist devices simultaneously. This is your demo-ready milestone.
Design simple PCB (EasyEDA, free). 3D-print wrist enclosure with silicone strap. Add LiPo charging. Test 10 simultaneous ESP-NOW devices to one gateway — validate no packet loss. Build battery life optimization: 10-second active sensing, 2-second deep sleep between samples. Target: 10-hour battery life. Implement offline buffering on gateway SD card.
Approach one MSME factory or construction contractor in Bhubaneswar/Odisha for a 4-week unpaid pilot. Deploy 10 devices. Have workers wear devices through full shifts. Collect real physiological data in Indian heat and physical labour conditions. Validate alert timing vs. worker-reported fatigue. This generates real data, real testimonials, and the first India-specific dataset for your model.
Write IEEE/INDICON conference paper on "Real-time physiological risk classification using TinyML fusion on ESP32 for Indian industrial workers." Submit to INDICON 2025 or NCC 2026. Build pitch deck with pilot data. File provisional patent on signal fusion + classification architecture. Present at Proxima / SOA Innovation Expo. Target: 1 paid POC contract (₹2–5L) from pilot client or referral.
Primary: IEEE INDICON (India Council International Conference) — directly relevant, national, reviewed. Deadline typically August for December conference.
Secondary: NCC (National Conference on Communications) — IIT-hosted, strong for signal processing papers.
International: IEEE EMBC (Engineering in Medicine & Biology Conference) — global platform for physiological sensing work.
Title: "VitalEdge: Real-Time Physiological Risk Classification for Industrial Heat Stress Prevention Using Multi-Modal Sensing and TinyML on ESP32"
Novel contributions:
The pilot data from Month 5 is what makes this publishable. Without real worker data, it's a lab paper. With it, it's a genuine contribution. The 10-worker pilot is not just business validation — it's your research dataset.
| Target | Type | Timeline | Angle |
|---|---|---|---|
| IEEE Sensors Journal | Q1 Journal | Month 12–18 | Multi-modal sensor fusion architecture |
| JMIR mHealth and uHealth | Q1 Journal | Month 14–18 | Occupational health monitoring outcomes data |
| Safety Science (Elsevier) | Q1 Journal | Month 18+ | Predictive incident prevention with field data |