TEJVONTechnologies Pvt. Ltd.
Back to Insights
AI · Offline-first 2 min readMay 2026

Our AI scanner failed at the one place it mattered most

The first version of our card scanner sent every photo to an AI model. It worked at our desks and failed on the expo floor.

T

Surendra — Tejvon

Senior Android & BLE Engineer

The first version of our business card scanner was simple. Take a photo, send it to an AI model, get clean contact data back. At our desks it worked well.

Then we thought about where the app is actually used. A trade expo hall is full of people on the same network. Signal drops between stalls. A sales rep has a few seconds between conversations to scan a card and move on. If the scan needs a network call, and the network is not there, the app has failed at the only moment it exists for.

So we changed the rule. The app must scan and extract a card with no network at all. AI became a bonus, not a dependency.

That meant moving to on-device OCR and accepting what comes with it. On-device text recognition is less forgiving than a large model. A slightly dark or blurred photo that a cloud model would still read comes back as noise. Accuracy was good on some cards and poor on others.

We did not fix that with a better model. We fixed it by making sure the model only ever sees a good photo. The app now runs a series of checks before and after capture, and across many iterations that took offline extraction to between 80 and 90 percent accuracy.

The lesson we took from it: decide where your product has to work before you decide how it works. For us the answer was a crowded hall with no signal, and every other decision followed from that.

Topics

AIOffline-first

Working on a similar challenge?

Let's discuss the architecture before you build.

BLE system design, OTA reliability, and connected product engineering — this is what TEJVON does every day.

Book a Technical Consultation