About us
We build AI that pauses, explains, and lets people decide.
UVERA (from Latin vera, “true”) is a trust layer for mobile money in Bangladesh, built by the UVERA team for AI DEV FEST 2026, Track 07.
Why
A scam transfer finishes in seconds. The people it hurts are often new to digital money, and nobody asks them to pause.
How
Seven AIs on a seeded synthetic world of 2.2 million events. Each one is compared with a simpler baseline and tested on people it never saw.
Principles
Never block. Say “not sure” when unsure. Show the evidence. A person approves every hold, and every number comes from a notebook run.
0
hours to build
0
Kaggle notebooks
0
AI components
0M
synthetic events
0
real customer records used
How it came together
Seventy-two hours, in order.
Day 1
Day 1
A world to learn from
We built a seeded synthetic world: 20,000 customers, 600 agents, 2,500 shops and 120 days of activity, with hidden scam processes, honest look-alikes, label noise and a leakage guard.
Day 2
Day 2
Seven AIs, nine notebooks
Every model ran on Kaggle against a baseline: grouped cross-validation, calibration, conformal “not sure”, exact TreeSHAP reasons, rolling forecasts and a grounded Gemini brief with a validator.
Day 3
Day 3
One website for everyone
Customer, agent and operations areas on a live API, a Trust Center that reads the notebook results, and this 3D field drawn from the same synthetic world.
7 Oct
7 Oct
The on-site final
A live demo, the report and the evidence, side by side.
The team
Three people, one weekend.
Tanvir Hossin
Team lead & design
Leads the team, the product story and the visual design.
Abdullah Rubab
Lead AI/ML & development
Machine learning, the website and the API: synthetic world, Pause Check, briefs and evaluation.
github.com/ABRUBAB
Abdur Rahman
Backend & AI/ML
Backend services and model notebooks.
Honest by default
What this is, and what it is not.
- All people, wallets, shops and messages are synthetic. Results show that the method works, not real-world accuracy.
- This is an independent hackathon prototype. It is not affiliated with, or endorsed by, any payment provider, and it describes no provider's internal problems.
- Every external dataset, model, API and component is listed with its license in the repository.
- Want the evidence? Open the Trust Center.