A model that scored well on a clean training set met real phone-camera photos from actual users and the failure rate jumped. The demo used good lighting and steady hands; production doesn't.
What actually breaks it
- Motion blur from a hand-held shot, especially in low light where exposure time increases.
- Uneven lighting — a face half in shadow, a document photographed under a single overhead bulb.
- Low resolution from an older or budget phone camera, or heavy compression from a messaging app.
- Skewed or rotated framing — a document photographed at an angle rather than scanned flat.
- Reflections and glare on glossy surfaces, ID cards and laminated documents especially.
The fix isn't a better model, it's better test data
Augment your training and evaluation data with exactly these degradations, deliberately, rather than only training on clean images. A model that's never seen blur will fail on blur no matter how good its architecture is.
See privacy by design for AI apps in India for a related production-readiness checklist.
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU Kanpur) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. Email pranjulrathour41@gmail.com.
