Authentico

• 

  Authentico Login/Welcome Page

Inspiration

It started with a story from an army mate who had been claiming Grab rides he never took — falsifying receipts to pocket the reimbursements. He got away with it for a while, until one day his boss had the bill of everything charged to his card. S$80,000. All employees were exploiting this loophole, which snowballed into a very big amount. TCaught red-handed, he had to pay everything back.

That story stuck with me. If it took paying — reviewing hundreds of claims in seconds, flagging anomalies that no finance team has the time or bandwidth to catch manually.

What surprised us was how few solutions exist in this space. Receipt fraud is a widespread, costly problem, yet the market is largely underserved. That gap is exactly why we decided to pursue this project.

What it does

Authentico is an AI-powered receipt authentication platform that helps finance teams detect fraudulent or tampered claims instantly. It operates on a two-tier system — admins (finance team) manage the platform and onboard staff members, who submit their receipts through their own accounts. Every submitted receipt is analyzed by a Vision Language Model that checks for signs of tampering like scratched-out values, altered amounts, and duplicate submissions. Results are surfaced with bounding boxes that visually pinpoint exactly where the anomaly was detected, giving finance teams clear, auditable proof rather than just a verdict. Authentico also automatically categorizes receipts into distinct types, reducing the manual sorting burden and laying the groundwork for automated payment workflows.

How we built it

We built Authentico using a Vision Language Model at its core, handling the visual forensic analysis of uploaded receipts. On top of that, we layered a two-tier role-based access system for admins and members, backed by an SMTP server for automated email notifications when new accounts are created. The bounding box detection and receipt categorization features were also powered by the VLM, allowing us to keep the architecture lean while maximizing what a single AI layer could do.

Challenges we ran into

The biggest challenge was prompt engineering the VLM to reason systematically through every forensic signal rather than jumping to a surface-level conclusion. Getting bounding boxes to accurately pinpoint tampered regions — rather than flagging the wrong areas — required significant iteration. Balancing the two-tier access system with a clean user experience under a tight deadline was also harder than expected.

Accomplishments that we're proud of

Coming into this hackathon, the four of us were fresh off our internships — AI was still relatively new territory for all of us. Building Authentico after hours, on top of our work commitments, in just a few days made this win feel very real. What we're most proud of isn't just shipping an MVP — it's how much we grew in the process. We came in with a basic understanding of AI and left with a genuine appreciation for what these tools are truly capable of, and more importantly, how to wield them to solve real problems.

What we learned

Building Authentico taught us how powerful Vision Language Models (VLMs) really are when applied to a focused, real-world problem. We went in knowing AI could read text — we came out appreciating how much it can see: inconsistent fonts, suspicious spacing, image artifacts, and signs of physical tampering that a human eye would easily miss after reviewing the hundredth receipt of the day. We also learned that combining multiple AI layers gives you far stronger signal than any single check alone. Metadata analysis catches what the VLM misses, and the VLM catches what metadata can't see. The two working together is what makes the verdict trustworthy.

On the technical side, prompt engineering for forensic analysis is a lot harder than it sounds — getting the model to reason systematically through every signal, rather than jumping to a conclusion, took more iteration than we expected.

Most importantly, we learned that AI is most valuable not when it replaces human judgement, but when it handles the volume so humans only have to step in where it truly matters.

What's next for Authentico

Future extensions to Authentico would include merging it with a payment system so that it can streamline the process of payment right after the human-in-the-loop has approved that the receipt claim is legitimate.

Built With

• css
• github
• next.js
• openai
• react
• react.js
• sql
• supabase
• typescript
• vercel
• vlm