I build things that tell you what they are actually sure of.
My name is Mohi Uddin. I build DebugAI on my own. The same habit shows up in everything I have worked on, in fields that have nothing to do with each other, and it is the reason this product looks the way it does.
DebugAI
Shipped
Every fix carries a verification state. Verified means a mechanical check ran and passed. Failed check means DebugAI disproved its own fix before showing it to you. Not verified means the confidence number is the model's own estimate and nothing tested it. Most tools quietly drop that third case into the same bucket as the first. It gets its own label here, because a confident wrong answer costs more than no answer.
OptiXon, health-sensing eyewear
Design complete, no physical prototype yet
A glasses temple arm carrying PPG heart rate, a 6-axis IMU, skin temperature and BLE on a 138 by 10.5mm four-layer flex-rigid board, with firmware that builds against the nRF5 SDK. The product analysis tags every single claim as verified, paper estimate, or not started, and the most useful file in the repository is a 51-entry bug log written as symptom, root cause, fix. Two of those entries were field-failure-class safety bugs caught before fabrication. Nothing has been physically built yet, which is exactly why the claims are labeled that way.
MatchLedger, bank reconciliation
Built and benchmarked
Matches ledger transactions against bank statements at scale. I replaced an embedding-based matcher with a rule-based engine: an exact hash join first, then a date-tolerance pass for clearing delays, then fuzzy matching on the small remainder only. On my own benchmark it reconciles 100,000 records against 98,000 in about four seconds, at a 95% match rate, with zero API calls. Every match carries a reason an auditor can read. The lesson I kept: use a model where it earns its place, and not one line further.
That is the whole thread. Separate what was checked from what was guessed, keep the record of what broke, and prefer a deterministic answer when one is available. DebugAI is that habit turned into a product.
GitHub · LinkedIn · mohi@debugai.io