Velosyti

Private AI for hospital records

Patient files carry a duty of privacy no cloud AI service should touch. Here is what private, on-site AI actually looks like.

· 3 min read · For hospital administrators and records teams

A hospital sits on years of patient files: scanned case sheets, discharge summaries, lab reports, sometimes recordings of consultations. Finding one old record can take a clerk half a day. AI can make that search take seconds, but sending patient data to a cloud service outside the country, or outside your control, is not something most hospitals can accept. Private AI is the alternative: the model runs inside your building, and nothing leaves.

Signs you need this

  • Staff spend real hours each week searching old case sheets or referral letters for one detail.
  • Your records are a mix of scanned paper, typed notes and Tamil, Hindi or English, and keyword search misses most of it.
  • You need to show, on demand, who asked for what patient information and when.

What a good system does

  • Runs fully offline, on a server inside your building or data centre, with no data sent to the internet.
  • Reads scanned handwriting and typed records together, in Tamil, Hindi and English.
  • Answers a question with the source page, not just a summary, so a doctor can verify it.
  • Logs every question and answer for audit, so access is traceable.

What it costs you

Expect the server and setup to be a one-time cost, followed by ongoing model updates and support. The bigger internal cost is discipline: deciding who can ask what, and making sure the audit log is actually reviewed, not just collected. A private system that nobody checks is not much safer than a cloud one. Assign the audit review to a specific person, on a specific schedule, rather than leaving it as a task everyone assumes someone else is doing, which is how audit logs quietly go unread for months.

Questions to ask any vendor

  • Does any data leave the building at any point, including for updates or support?
  • Can it read our scanned records as they actually are, not a cleaned-up sample?
  • Is every access logged, and can we produce that log for an audit?
  • What happens if the server goes down — do our staff lose search entirely, or is there a fallback?

How we can help

Kaappu is our models in a box, built for exactly this: courts, hospitals, banks and departments whose records cannot go to a cloud AI. It runs fully offline and every answer is logged with its source. See how it works — explore Kaappu — or talk to us.

References

  1. Ayushman Bharat Digital Mission
  2. MeitY — data protection framework