Building a face recognition system changes how you think about data. When the thing you might leak is someone's face, "we take privacy seriously" stops being a sentence on a page and becomes a set of architectural decisions. These are the ones I now make by default, on every AI product, not just FaceVision.
Minimise what you collect
Ask what the feature needs, not what might be useful later. A verification feature needs an embedding; it does not need the photo. A transcription feature needs the text; it rarely needs to keep the audio. Every field you do not store is a field you cannot leak, cannot be asked for, and do not have to delete.
Prefer on-device inference
If a model can run in the browser or on the phone, run it there. Raw images and audio never cross the network, and the privacy claim becomes verifiable by anyone with developer tools open. ONNX Runtime Web made this practical for vision; the same logic applies to small speech and text models.
Store derived data, not source data
Embeddings instead of faces. Extracted fields instead of scanned documents where the workflow allows. Chunks and citations instead of full copies where you can regenerate them. Derived data is smaller, less sensitive, and easier to justify.
Retention and deletion as features
- Every table with personal data has a retention rule and a job that enforces it.
- Deletion is one operation that removes the user everywhere — including vector stores and caches, which teams forget.
- Logs never contain raw inputs; they contain ids and hashes.
Consent that says what happens
"We use AI to improve your experience" is not consent. "We compute a numeric face signature in your browser and store only that signature to verify you next time; we never store your photo" is. India's Digital Personal Data Protection Act pushes in this direction — purpose limitation, notice, erasure — and building for it now is cheaper than retrofitting later.
Privacy by design is not a compliance chore. It is the reason a college, a clinic or a client will let your AI product anywhere near their people.
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.
