Bharat Harshan - a 28 year old Zomato rider from Kurla, Mumbai.
It's Tuesday, 6 PM. Peak delivery hour. Then the monsoon hits. In 20 minutes, orders in his zone crash by 70%. Customers pay surge pricing. The platform keeps earning. Ravi ends the night with barely ₹90.He didn’t fail to work. The weather failed him. That’s why Jugaad exists.
Rainfall crosses 25 mm/hr. The trigger fires. GPS verifies his zone. ₹680 lands in his UPI. No forms. No claims. No waiting.
Because insurance shouldn’t begin after the crisis. It should respond the moment the crisis starts.
- What is Jugaad?
- The 7 Scenarios We Cover
- System Architecture
- ML & AI Engine
- Zero-Trust Security
- Premium & Payout Model
- IRDAI Compliance Checklist
- Legal & Regulatory
- Tech Stack
- Running Locally
- Deployed App
- Meet the team
Jugaad is a fully automated parametric income insurance platform for India's 5–8 million food delivery workers.
When weather stops deliveries, their income shouldn't stop too.
Traditional insurance requires filing claims, waiting weeks, proving loss. Jugaad doesn't. An objective weather threshold is crossed → GPS verifies the rider is in the affected zone → ₹ lands in their UPI wallet. Under 2 hours. Zero paperwork. Zero human intervention.
External disruption fires → GPS verified in real-time → ₹ transferred via UPI in < 2 hrs
(objective API) (zero-trust engine) (Razorpay sandbox)
Parametric insurance pays on a measurable event — not on proving loss. No forms. No agents. No waiting. The algorithm pays — instantly, fairly, automatically.
| Pain Point | The Reality |
|---|---|
| 💸 Fixed costs never pause | Bike EMI ₹2,800–3,200/month + fuel ₹180–220/day — due whether they work or not |
| 🎯 The incentive trap | One 2-hr rain event → miss weekly slab → lose ₹800 bonus on top of lost orders |
| 90% have zero savings buffer; 68% spend more than they earn | |
| 🌧️ Platforms profit, riders don't | Zomato charges 40% surge in rain — riders receive ₹0 extra |
Every trigger uses objective, third-party, government-verified data. No forms. No calls.
| # | Scenario | Trigger | Payout | Data Source |
|---|---|---|---|---|
| 🌡️ 1 | Extreme Heat | IMD Red Alert + heat index ≥44°C for 2+ hrs | 70% avg daily earnings | IMD + Open-Meteo |
| 🌧️ 2 | Heavy Rain & Flooding | Rainfall ≥25mm/hr for 90+ min in rider's zone | 70% shift earnings + Slab Shield | IMD + Open-Meteo |
| 🌫️ 3 | Dense Winter Fog | IMD Dense Fog advisory + visibility <100m for 3+ hrs | Half-shift income (7–11 PM) | IMD Fog Bulletins |
| 😷 4 | Severe AQI Pollution | CPCB AQI ≥400 for 3+ consecutive hours | Health Safety Supplement | CPCB API (free, hourly) |
| 🚫 5 | Bandh / Curfew | NewsAPI 3+ sources confirmed + zone orders drop ≥65% | Full declared shift income | GNews API + platform mock |
| 🚧 6 | Metro / Congestion | Avg travel time >200% of historical mean | Per-hour Congestion Supplement | OSRM (open-source) |
| 💀 7 | Platform App Outage | Downdetector 500+ reports + outage >45 min peak hours | Hourly rate × outage duration | Downdetector (independent) |
The Jugaad Principle: Every payout is triggered by objective, third-party, government-verified or publicly monitored data. The worker never has to prove anything. The insurer never has to investigate anything. The algorithm pays — instantly, fairly, automatically.
⚖️ Coverage Scope & The Golden Rules
- ✅ Income Loss Protection Only: This platform is strictly designed as a safety net for lost hours and unearned wages due to external events.
- 🚫 Exclusions: We strictly exclude any coverage for health issues, life insurance, accidents, or vehicle repairs.
| Endpoint | Purpose |
|---|---|
/api/users/active |
Returns live active workers with location |
/api/claims/trigger |
Receives validated triggers and records payouts |
/api/admin/metrics |
Dashboard stats + fraud queue |
POST /api/claims/trigger
Content-Type: application/json
{
"user_id": "worker_123",
"scenario": "heavy_rain",
"confidence": 0.91,
"payout": 680
}To enable precise, ward-level trigger validation, we implemented native spatial queries using PostGIS.
-- Convert raw coordinates into geographic points
SELECT ST_SetSRID(ST_MakePoint(longitude, latitude), 4326);
-- Find workers within 500m of trigger zone
SELECT *
FROM workers
WHERE ST_DWithin(
location,
ST_SetSRID(ST_MakePoint(:lon, :lat), 4326),
500
);
Jugaad's intelligence layer does three things: price risk fairly each week, predict disruptions before they hit, and validate every trigger before a payout fires.
Every data source used is free, open, or government-published — zero paid API dependency.
┌─────────────────────────────────────────────────────────────────────────────┐
│ ML PIPELINE (runs every Sunday) │
│ │
│ 1. Earnings DNA Profiler → Gradient Boosted Regressor │
│ Personalised income fingerprint per rider at onboarding │
│ Powers Slab Shield: predicts if disruption kills weekly bonus │
│ │
│ 2. Dynamic Premium Engine → XGBoost │
│ 11 risk factors → weekly risk score (0–1) → premium ₹29–₹99 │
│ Runs Sunday midnight · Premium confirmed in DB by 6 AM │
│ │
│ 3. Risk Zone Classifier → Random Forest │
│ Labels every delivery zone: LOW / MEDIUM / HIGH per scenario │
│ Kurla ≠ Powai — same city, different zone, different premium │
│ │
│ 4. Disruption Predictor → Prophet │
│ Forecasts P(trigger) per scenario per zone for next 7 days │
│ P > 0.40 → seasonal multiplier rises + Sunday rider alert fires │
│ │
│ 5. Claim Trigger Validator → Rule engine + ML hybrid │
│ Hard threshold check → dual-source confirm → confidence score │
│ ≥ 0.85 → auto-pay · < 0.85 → human review queue │
│ │
│ 6. Continuous Learning → Weekly retraining loop │
│ Actual loss / Expected loss → nudges all models ±15% per week │
│ Basis risk shrinks every monsoon, every fog winter, every heatwave │
└─────────────────────────────────────────────────────────────────────────────┘
14-day monsoon stress test
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Premium inflow → tracked daily across simulated user pool
Payout outflow → weather-driven triggers, zone-by-zone
Fraud losses → ML-flagged + blocked by kinematic engine
Calibrated BCR = 0.67
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✓ Sustainable band: 0.6 – 0.7
✓ Pool remains solvent under 14-day continuous extreme stress
✓ Not exploitative (< 0.8), not loss-making (> 0.5)
"GPS alone is dead. We require five independent signal layers — all of which must agree before a rupee leaves the pool."
Every session: live HTML5 device coordinates are sent — not stored, not cached, not user-typed. Declared city is geocoded via Nominatim. Distance delta triggers lockout if violated.
User types "Bangalore" at onboarding
↓
Nominatim geocodes → 12.97°N, 77.59°E
↓
Device GPS reads → 8.52°N, 76.93°E ← actual location: Trivandrum
↓
Distance delta: 743 km
↓
🔴 KINEMATIC VIOLATION — Security Lockout fires on dashboard entry
This is live and demoed. Type Bangalore. Be in Trivandrum. Watch the red screen fire.
Layer 1 — Device Trust
Mock-location flags · rooted/tampered device detection · Android Play Integrity
If the device environment is compromised → GPS output rejected outright
Layer 2 — Mobility Authenticity
Last 15–30 min GPS telemetry analyzed · teleport jumps flagged
Route matched against real road network via OSRM
Accelerometer/gyroscope used as soft motion corroboration
Layer 3 — Operational Eligibility
Active delivery session required · recent order acceptance checked
Last trusted platform event timestamp + location verified
Worker at home with no active trip → does not qualify, regardless of GPS
Layer 4 — Ring Detection (Sybil Defense)
500 workers becoming eligible in same short window → automatic flag
Shared device fingerprints + shared UPI accounts → graph detection
Synchronized teleport-into-zone patterns → cluster quarantine
If real flood: real cancellation spike, real merchant impact exist.
A syndicate has none of these. That absence is the signal.
| Tier | Condition | Action | Worker sees |
|---|---|---|---|
| ✅ Auto Approve | All signals consistent. Device clean. Route plausible. Active trip confirmed. | Payout fires < 2 min | "₹680 credited. Stay safe." |
| ⏳ Soft Hold | Weak GPS in heavy rain, brief network drop, delayed sync | Wait 10–15 min for signals to stabilize | "Payout being verified. Usually resolves in 15 min." |
| 🔴 Quarantine | Low device trust + impossible route + no active trip + ring-risk | Held → human review | "Under quick review. We'll update you shortly." |
The Soft Hold tier exists because honest workers in heavy rain will have degraded signals. Missing data ≠ fraud. Only contradictory evidence escalates.
To ensure uninterrupted service during external API failures:
-
If Open-Meteo / CPCB APIs fail: → Use last known valid data snapshot (cached in Redis)
-
If API latency exceeds threshold: → Trigger retry with exponential backoff
-
If all sources fail: → System enters grace mode → Critical payouts (SOS / high confidence cases) still processed
if (!weatherAPI.available()) {
data = cache.get("last_valid_weather");
}A risk-adjusted, income-based micro-insurance framework. Premiums computed every Sunday, debited every Monday at 9 AM via UPI auto-mandate.
Step 1 — Base Premium
Base = Weekly Income × Base Rate (1.5% – 2%)
ML cross-checks declared income against zone average to prevent fraud at source
Step 2 — Risk Multipliers
Adjusted = Base × City Risk (0.85–1.40×)
× Shift Factor (0.85–1.20×)
× Platform Factor (1.00–1.10×)
× Zone Factor (0.80–1.30×)
Step 3 — Affordability Cap
Final = min(max(Adjusted, ₹29), ₹99)
Hard floor ₹29 — always affordable even for part-time workers
Hard ceiling ₹99 — never more than ~2.5% of full-time weekly income
| Tier | Weekly Premium | Coverage | Triggers | Max Weekly Payout |
|---|---|---|---|---|
| 🟡 Basic | ₹29 – ₹49 | 60% daily income | Any 3 of 7 | ~₹720 |
| 🔵 Standard ✦ | ₹49 – ₹79 | 70% daily income | All 7 + Slab Shield | ~₹1,400 |
| 🔴 Full Shield | ₹79 – ₹99 | 80% daily income | All 7 + Platform Bridge | ~₹2,400 |
New Premium = Old Premium × (Actual Loss / Expected Loss)
Change capped at ±15% per week — no premium shock
A dry November week costs less.
A monsoon July week costs more.
Workers see exactly why, every Sunday morning.
| # | Requirement | Status | Implementation |
|---|---|---|---|
| 1 | Trigger objective & verifiable | ✅ | AQI >300 / rain ≥25mm/hr from CPCB + IMD — quantifiable, public, tamper-proof |
| 2 | Health, life & vehicle excluded | ✅ | Legal disclaimer hardcoded in UI: "Covers income loss from weather only" |
| 3 | Payout automatic | ✅ | Trigger → GPS verify → UPI transfer within 2 hours via cron |
| 4 | Pool financially sustainable | ✅ | BCR 0.67 — 14-day monsoon stress-tested, PDF artifact in /ml/reports/ |
| 5 | Fraud detection on data, not behaviour | ✅ | GPS × login cross-check, kinematic ML, Sybil ring detection |
| 6 | Premium collection frictionless | ✅ | ₹29–₹99/week via UPI auto-mandate — zero manual steps |
| 7 | Dynamic pricing, not flat | ✅ | XGBoost: HIGH risk → +10% premium, LOW risk → discount, seasonal + zone adjustments |
| 8 | Adverse selection blocked | ✅ | 48-hour enrollment lockout before weather red alerts (FastAPI hard block) |
| 9 | Operational cost near zero | ✅ | Fully containerised straight-through processing, zero human in claims loop |
| 10 | Basis risk minimized | ✅ | Ward-level GPS mapping via Nominatim — not broad 15km city radius |
The SS Code formally recognises gig workers as eligible for welfare benefits for the first time in India.
GigShield implements the 90/120-day engagement rule in backend math, not just UI:
Single platform worker → 90 active days in FY → state-backed social security unlocked
Multi-apping worker → 120 active days in FY → state-backed social security unlocked
Progress bar in-app switches targets automatically based on platform sync data.
Three data types collected. Each with an explicit, un-pre-checked consent box:
| Data | Purpose | Consent Requirement | Built? |
|---|---|---|---|
| 📍 Continuous GPS | Verify worker is in trigger zone | Separate consent screen | ✅ |
| 🏦 Bank / UPI account | Payout disbursement | Explicit consent + KYC mention | ✅ |
| 📱 Platform activity | Confirm active delivery days | Data sharing agreement | ✅ |
Storms kill internet. Workers can't lose coverage when the network drops.
Worker loses internet during storm
↓
App detects network loss instantly
↓
Red fallback banner appears on screen
↓
SMS auto-populated with exact GPS coordinates
↓
POST /api/webhook/sms-sos (Twilio-compatible endpoint)
↓
Emergency ₹500 claim processed
Bypasses 2-minute cron timer entirely
No internet. Still covered. No action needed from the worker.
| Technology | Purpose |
|---|---|
| React.js + Vite | Component-based PWA with fast builds |
| Tailwind CSS | Responsive, mobile-first styling |
| Framer Motion | Smooth UI transitions and animations |
| Firebase Auth (Phone OTP) | Secure, frictionless onboarding |
| Socket.io client | Real-time payout notifications |
| Technology | Purpose |
|---|---|
| Node.js + Express | Core API orchestration + policy CRUD |
| Socket.io | Real-time frontend push (payout alerts) |
| node-cron | Automated trigger engine (every 2 min) |
| FastAPI (Python) | ML model serving + inference endpoints |
| JWT | Secure session handling |
| Model | Purpose |
|---|---|
| XGBoost | Dynamic premium calculation (11 risk factors) |
| Prophet | Disruption probability forecasting (7-day) |
| Isolation Forest | Anomaly + fraud detection |
| LightGBM | Fraud scoring (tabular, handles missing data) |
| scikit-learn | Pipelines + feature engineering |
| MLflow | Experiment tracking + model versioning |
| Source | Data |
|---|---|
| Open-Meteo | Rain, heat index, fog, wind |
| CPCB / OpenAQ | AQI (hourly, city-specific) |
| IMD | Red alerts, fog bulletins, rainfall bulletins |
| GNews API | Bandh / curfew keyword detection |
| OSRM | Open-source travel time routing |
| Downdetector | Platform outage signals (independent) |
| Nominatim | Geocoding for ward-level location mapping |
| Tool | Purpose |
|---|---|
| PostgreSQL + PostGIS | Structured storage with geographic coordinate support |
| Redis | Real-time trigger caching + fast lookups |
| Docker + Docker Compose | Full multi-container orchestration |
| GitHub Actions | CI/CD pipeline |
| Razorpay Sandbox | Instant payout simulation |
| Twilio | SMS SOS webhook |
Worker Dashboard: https://jugaad-frontend-git-main-bhavanaharshans-projects.vercel.app/
Admin Dashboard (Password: admin123): https://devtrails-1-40gq.onrender.com/
- Docker + Docker Compose installed
- Node.js 18+
- Python 3.10+
- Git
# 1. Clone the repository
git clone https://github.com/bhavanaharshan/DEVTrails.git
cd gigshield
# 2. Copy environment config
cp .env.example .env
# 3. Boot the entire stack
docker-compose up --build| Member | Domain | Key Contributions |
|---|---|---|
| Priya | ML / Data Science | BCR simulation, XGBoost pricing engine, kinematic fraud model, monsoon stress test, adverse selection lockout API |
| Samridhi | Geo / Backend Orchestration | Ward-level GPS mapping, Nominatim integration, trigger cron orchestrator, fraud routing pipeline |
| Neema | Backend / Infrastructure | Express API, PostGIS migrations, Docker Compose, UPI auto-deduction architecture, admin hub |
| Bhavana | Frontend / Compliance | DPDP consent flow, kinematic lockout UI, real-time dashboard, offline SOS, WebSocket notifications |
| Deliverable | Status |
|---|---|
| ✅ 5-minute demo video | Watch here |
| ✅ Pitch deck (PDF) | [Click here] |
| ✅ Source code (default branch) | This repository |
| ✅ Docker Compose (local setup) | docker-compose up --build |
Built with ❤️ for India's 8 crore invisible workers
"He kept riding next week. His family is okay. That's what insurance is really for."
