Inspiration
While researching India's priority sector lending crisis, we discovered that 16% of all government loans had turned into NPAs by March 2025 — not because borrowers couldn't repay, but because nobody verified whether the money was used correctly in the first place. A bank officer physically visiting a village in Haryana to check if a farmer bought a tractor costs ₹2,000 and takes 3 weeks to schedule. Multiply that across lakhs of beneficiaries and the system simply collapses. We asked one question — why is a government officer driving hours to look at a tractor when every beneficiary already has a smartphone with a camera and GPS? Loan Lens was born from that question.
What it does
Loan Lens is a zero-visit, AI-powered loan utilization verification platform with two interfaces working together.The beneficiary opens the mobile app, logs in via Aadhaar-linked OTP, and photographs their purchased asset — tractor, seeds, irrigation pump, livestock, whatever the loan was sanctioned for. The moment the photo is captured, GPS coordinates and a tamper-proof timestamp are automatically embedded. An on-device AI model instantly analyzes whether the asset is genuinely present, checks for duplicate or recycled images, and produces a confidence score. If there's no internet, everything saves locally and syncs automatically the moment connectivity returns.The bank officer opens the web dashboard from anywhere, sees the full verification queue, and for each case gets the photo, a GPS map pin, the timestamp, and the AI confidence score — all in one view. They approve or reject with a single tap. The beneficiary is notified instantly on their app.Verification that used to take 2–4 weeks now takes 2–4 hours.
How we built it
We built the Loan Lens prototype as a single-file HTML/CSS/JavaScript application that fully simulates both the beneficiary mobile experience and the officer dashboard — no backend required for the demo.
Challenges we ran into
Offline-first architecture was harder than expected. Making the app work seamlessly without internet — capturing, queuing, encrypting, and then silently syncing — required careful handling of Hive local storage, Firestore offline persistence, and connectivity listeners all working together without conflicts. GPS spoofing edge cases. We realized early that a fraudster could simply fake their GPS location. We had to build a secondary check that cross-references the submitted coordinates against the beneficiary's registered address in Firestore and flags anything beyond 1km. On-device AI without a custom trained model. Training a domain-specific model for Indian agricultural assets from scratch in 7 days was impossible. We used a pre-trained MobileNet base and adapted the inference pipeline to work within the constraints of low-end Android devices common in rural India. Balancing UI simplicity for low-literacy users. The beneficiary app needed to work for someone who has never used a fintech app before. Every screen went through multiple iterations to reduce steps, increase button sizes, and make the flow completely self-explanatory without reading anything.
Accomplishments that we're proud of
We are proud that we built a fully functional, end-to-end working prototype in 7 days — not a mockup, not a slideshow, but a live system with real camera capture, real GPS tagging, real Firebase sync, and a working officer dashboard. We are proud of the offline-first architecture that actually works in zero-connectivity conditions — because that's the exact reality of the villages this product is meant for. We are proud that the entire system costs under ₹2 per beneficiary to operate at scale, compared to ₹2,000 for a physical field visit — a 1000x cost reduction. And we are proud that this solves a problem that is real, documented, and affecting crores of people right now — not a hypothetical future problem.
What we learned
We learned that the hardest part of building for rural India is not the technology — it's the assumptions you have to unlearn. We kept designing for users like ourselves until we forced ourselves to think about someone with a 2G connection, a 4-year-old Android phone, and low digital literacy. Every technical decision after that became clearer. We learned that offline-first is not a feature you add at the end — it is a foundational architecture decision that affects every other layer of the system. Building it in late would have meant rebuilding everything. We learned that AI on the edge — running inference on the device itself rather than sending images to a server — is not just faster, it is more trustworthy, more private, and more resilient. For a government compliance use case, that matters enormously. We also learned how to move fast as a three-person team under pressure, divide ownership clearly, and make decisions with incomplete information — which is probably the most valuable thing we took away from this hackathon.
What's next for Loan Lens
Short term — Pilot in Haryana. We have designed the system to be pilot-ready in 30 days. We want to partner with one regional rural bank or NABARD district office in Haryana, onboard 500 beneficiaries, and generate real verification data to prove the model works outside a hackathon environment. SMS/USSD fallback. Not every beneficiary has a smartphone. We will build a feature-phone compatible submission flow using SMS and USSD codes so the system works for 100% of borrowers, not just smartphone users. Voice prompt support. For low-literacy users, we will add voice instructions in Hindi and regional languages that guide the beneficiary through the capture process step by step without requiring them to read anything. Custom AI model training. We will collect real geo-tagged asset images from field pilots and train a domain-specific model for Indian agricultural and small business assets — improving accuracy well beyond the current MobileNet base. Integration with government DBT infrastructure. The long-term vision is direct integration with the Direct Benefit Transfer portal so that loan disbursement, utilization verification, and repayment tracking all live in one auditable, transparent pipeline — reducing NPA at the source rather than chasing it after the fact.
Log in or sign up for Devpost to join the conversation.