DeliverGuard AI is a parametric micro-insurance platform designed for delivery partners of Zomato.
It protects workers from income loss caused by external disruptions such as rain, traffic, extreme heat, and environmental conditions.
The system uses AI monitoring, OCR verification, and fraud detection to ensure fair and automated payouts.
Problem Statement
Delivery workers depend on daily or weekly earnings, but external factors such as:
- Heavy rain
- Flood
- Traffic congestion
- Extreme heat
- Environmental disturbances
- Curfew / strike
can reduce or completely stop their ability to work.
Currently, there is no reliable system to compensate short-term income loss.
Deliverable Expectations and it's Solutions
| Expectations | Solutions |
|---|---|
| Onboarding | OCR-based income extraction with simple user profiling |
| Risk Profiling | AI-powered analysis using weather, AQI, traffic, and behavioral data |
| Policy Creation | Weekly income-based pricing with dynamic risk evaluation |
| Claim Triggering | Automated detection of disruptions (rain, AQI, traffic) |
| Payout Processing | Secure and instant bank transfers |
| Analytics Dashboard | Real-time insights on claims, payouts, and risk trends |
| Fraud Detection | GPS, IP tracking, device fingerprinting, and behavior analysis |
Phase 1 – Building a Secure & Reliable Foundation
In the first phase of our project, we focused heavily on trust, verification, and fraud prevention, which are critical for any insurance system.
Adversarial Defense & Anti-Spoofing
DeliverGuard AI implements a multi-layer fraud detection system that validates user activity using location, device, network, and behavioral signals. Each feature is designed to detect a specific type of fraud and contribute to a unified risk score.
Detection Layer
1. GPS Verification System
Problem:
Users can spoof GPS using fake location apps, making it appear they are working when they are not.
Solution:
Continuously validate location consistency instead of trusting a single GPS point.
How it Works:
- Collect GPS coordinates periodically (every few seconds)
- Store previous and current locations
- Calculate distance between points
- Detect abnormal jumps (e.g., 100 km in seconds)
- Flag inconsistent movement
Tech Stack:
-
navigator.geolocation - Haversine formula
2. IP Address Verification
Problem:
User’s network location may not match their physical location.
Solution:
Cross-check IP-based location with GPS coordinates.
How it Works:
- Extract IP address from request
- Use IP geolocation API to get location
- Compare IP location with GPS location
- Calculate distance mismatch
- Flag large inconsistencies
Tech Stack:
- ipapi / ipinfo
3. VPN Detection Mechanism
Problem:
Users can hide their real location using VPN or proxy services.
Solution:
Detect anonymized IP addresses and unusual location switching.
How it Works:
- Check IP against VPN/proxy database
- Detect rapid country switching
- Identify high-risk IP patterns
- Mark suspicious sessions
Tech Stack:
- ipqualityscore
- proxycheck.io
4. Device & Emulator Detection
Problem:
Fraudsters create multiple fake accounts using emulators.
Solution:
Generate unique device fingerprint and detect emulators.
How it Works:
- Collect device information (OS, browser, screen)
- Generate hashed deviceId (SHA-256)
- Detect emulator signatures
- Track multiple accounts on same device
Tech Stack:
- FingerprintJS
- Crypto hashing
5. Movement & Speed Analysis
Problem:
Fake GPS creates unrealistic movement patterns.
Solution:
Analyze speed and movement consistency.
How it Works:
- Calculate distance between GPS points
- Compute speed = distance / time
- Detect:
- No movement
- Unrealistic speed (>120 km/h)
- Sudden jumps
- Flag suspicious behavior
Tech Stack:
- Haversine formula
6. Route Validation System
Problem:
Fake routes do not follow real-world roads.
Solution:
Compare user path with actual map routes.
How it Works:
- Track sequence of GPS points
- Map points onto real road network
- Detect invalid paths (through buildings/water)
- Validate route realism
Tech Stack:
- OpenStreetMap
- Leaflet
7. Log-Based Monitoring System
Problem:
Fraud patterns cannot be identified from a single event.
Solution:
Maintain historical logs for analysis.
How it Works:
- Store all tracking data (GPS, IP, device, timestamp)
- Analyze repeated anomalies
- Detect long-term suspicious patterns
- Flag repeat offenders
Tech Stack:
- MongoDB
- Logging system
Intelligence Layer
8. Spatio-Temporal Correlation
Problem:
Fraudsters operate in coordinated groups.
Solution:
Analyze location and time relationships.
How it Works:
- Compare multiple users’ activity
- Identify same location + same time patterns
- Detect clustering behavior
- Flag coordinated activity
Tech Stack:
- MongoDB aggregation
9. Shared IP & Device Detection
Problem:
One attacker controls multiple accounts.
Solution:
Detect shared device and IP usage.
How it Works:
- Store deviceId and IP for each user
- Group users with same identifiers
- Detect abnormal sharing
- Flag accounts for investigation
Tech Stack:
- Backend grouping logic
10. Fraud Ring Detection
Problem:
Large-scale fraud networks operate together.
Solution:
Identify clusters of users with similar behavior.
How it Works:
- Analyze user patterns (routes, timing, devices)
- Detect repeated similarities across accounts
- Build clusters of related users
- Identify fraud networks
Tech Stack:
- Graph-based analysis
11. Behavioral Pattern Analysis
Problem:
Fake users behave unnaturally compared to real users.
Solution:
Analyze behavioral patterns over time.
How it Works:
- Track delivery frequency and timing
- Identify unusual consistency
- Detect robotic or scripted behavior
- Flag anomalies
Tech Stack:
- Statistical models
12. Multi-Signal Fusion Engine
Problem:
Single signal is unreliable.
Solution:
Combine all signals for stronger detection.
How it Works:
- Collect signals (GPS, IP, device, behavior)
- Assign weight to each signal
- Combine into unified decision
- Detect fraud based on multiple indicators
Tech Stack:
- Rule-based engine
Risk Scoring System
13. Dynamic Risk Scoring
Problem:
Not all anomalies indicate fraud.
Solution:
Assign weighted risk scores.
How it Works:
- Each anomaly adds risk points
- Combine scores from all modules
- Calculate final risk score
- Classify user risk level
Tech Stack:
- Rule-based scoring system
Final Insight
DeliverGuard AI combines detection, intelligence, and risk scoring layers to build a robust, real-time fraud prevention system that ensures security while minimizing false positives.
OCR-Based Income Verification
To ensure accurate and automated income verification, DeliverGuard AI uses OCR (Optical Character Recognition) powered by EasyOCR.
Why OCR is Used
- Eliminates manual verification of income proof
- Automatically extracts data from transaction screenshots
- Speeds up onboarding and claim validation
- Reduces human errors and improves efficiency
How It Works
- The user uploads a screenshot of their transaction or earnings proof
- The system uses EasyOCR to extract text from the image
- Key details such as:
- Platform name (e.g., Zomato, Swiggy)
- Transaction amount
- Date and time
- Extracted data is processed and structured in the system
Example:
Input:
INR 3139 credited via ZOMATO
Output:
Premium Rate : 8% (Standard Plan)
Insurance Premium : ₹251
Fraud Prevention Mechanism
OCR alone cannot verify whether an image is original or edited. Therefore, DeliverGuard AI combines OCR with multiple validation techniques:
- Identifies inconsistent formatting in manipulated screenshots
- Flags duplicate or reused images
- Uses metadata analysis as an additional validation layer
Cross-Verification with Delivery Platforms
To enhance reliability, the system can cross-verify user income with Zomato delivery platforms:
- The extracted income data is compared with actual earnings records
- Ensures that the submitted screenshot matches real transaction history
- Prevents fraud caused by edited or AI-generated screenshots
- Acts as a strong validation layer beyond OCR
Since OCR only reads visible text, cross-verification ensures authenticity by validating the data from trusted sources.
Workflow
Technology Used
- EasyOCR – for text extraction from images
- Image preprocessing – to improve OCR accuracy
- Backend validation logic – for data matching
- Platform verification (Zomato integration) – for authenticity checks
System Workflow
- Collect user data
- Validate inputs
- Analyze patterns
- Assign risk score
- Trigger actions
Phase 1 Outcome
By the end of Phase 1, we successfully built:
- A robust fraud detection system
- A secure onboarding and verification pipeline
- A trust-first foundation for insurance processing
This allowed us to confidently move to the next stage.
Phase 2 – Expanding Intelligence & Insurance Logic
After establishing a strong foundation, we progressed to Phase 2, where we focused on making the system smarter and more user-centric.
Insurance Plans
| Plan | Premium | Hour Threshold | Max Weekly Payout |
|---|---|---|---|
| Basic | 5% | 8 hrs | ₹2000 |
| Standard | 8% | 6 hrs | ₹4000 |
| Premium | 10% | 4 hrs | ₹8000 |
Premium Calculation
Weekly Premium = Weekly Income × Plan %
Example (Weekly Income = ₹7000):
- Basic → ₹350
- Standard → ₹560
- Premium → ₹700
Payout Calculation
Hourly Income
Weekly Income ÷ 42 (6 hours/day × 7 days)
Claim Types
- Day Claim → 6 × Hourly Income
- Hour Claim → Threshold × Hourly Income
Final Rule
Final Payout = min(calculated amount, plan limit)
Disruption Detection
The system uses real-time APIs:
- Weather API → Rain / flood
- AQI API → Pollution
- Traffic API → Congestion
Trigger conditions:
- Rainfall ≥ 50 mm/hr
- AQI ≥ 300
- Traffic ≥ defined threshold
Loyalty Rewards
Levels
- Silver → 3 months
- Gold → 6 months
- Diamond → 1 year
Benefits
- Fuel rewards
- Premium discounts
- Faster claim processing
- Increased coverage
Final Thought
DeliverGuard AI is not just an insurance system — it is a smart protection layer for gig workers, ensuring they are supported even when they cannot work.
By combining verification, intelligence, and automation, we are building a system that is:
- Fair
- Secure
- Reliable
- Worker-centric
Top comments (0)