Product story
AI Engineering · 2 min read
Offline AI card scanning engine
On-device business card extraction that works with no network, at 80 to 90 percent accuracy, built from scratch over many iterations.
- Client
- Tejvon product: AI Lead Capture
- Industry
- Sales technology
- Capabilities
- AI Engineering, Mobile Apps
- Status
- Beta
Highlights
- 80 to 90% extraction accuracy with zero network
- No capture button: the app takes the photo when it is good enough
The challenge
The first version sent each card photo to a cloud AI model. It worked at a desk and failed on the expo floor, where networks are congested and sales reps have seconds between conversations.
Our solution
- 01Made the app offline-first: on-device OCR does the core job and cloud AI only improves results when a connection exists.
- 02Fixed accuracy by guaranteeing a good photo rather than a better model: 12 to 15 checks run around every capture.
- 03Moved every decision from the user to the app: the flash turns on in low light and the photo is taken the moment the frame is sharp.
- 04Guided the flow from front to back of the card, and cropped automatically with edges the user can adjust.
What we delivered
- On-device OCR extraction pipeline
- Capture quality checks with automatic flash and automatic capture
- Front and back guidance
- Automatic crop with adjustable edges
Impact
80 to 90%extraction accuracy with zero network
No capture button: the app takes the photo when it is good enough
Technology
KotlinJetpack ComposeCameraXML KitRoomHilt
Have a problem like this?
Start with a free 1-hour architecture review with a senior engineer.