Teaching AI to read a doctor's handwriting

"Doctor's handwriting" is a punchline for a reason. But there's a real engineering problem hiding inside the joke — and it's much harder in India than the textbook version of handwriting recognition you'll find in a research paper.

Why Indian prescriptions are hard

• Languages mix mid-line — a drug name in English, a dosage instruction in Hindi or Tamil, a note in shorthand.
• Abbreviations are everywhere and aren't standardised: 'OD', 'BD', 'HS', 'x5d', '1-0-1' all carry meaning a model has to learn.
• Drug names are long, similar-looking, and easy to confuse — a misread can be dangerous, not just wrong.
• Every doctor writes differently, and the same doctor writes differently when they're in a hurry.

Intelligent Character Recognition (ICR) is the umbrella term for turning handwriting into structured text. Off-the-shelf OCR isn't built for this. It expects printed characters, a single language, and clean lines. A prescription gives it none of those.

Capturing strokes, not just pixels

Because the Rx-01 pad captures writing as it happens, we don't only get a picture of the finished prescription — we get the strokes: the order, direction, and timing of each pen movement. That signal is enormously helpful. Two letters that look identical as static images often look very different as a sequence of movements. Starting from strokes rather than pixels gives the model a head start.

Why safety comes before accuracy

It's tempting to chase a single accuracy number. We think that framing is misleading on its own. A model that's right most of the time but confidently wrong on a look-alike drug name is worse than one that knows when to ask. So we care less about a headline percentage and more about calibrated confidence — the system should know what it doesn't know.

The goal isn't a model that's always right. It's a model that's never confidently wrong about something dangerous.

When confidence is low, Adicare doesn't guess silently. It flags the line for a quick confirmation. When it's reading a drug-and-dose pair, it cross-checks against known interactions and dosing ranges before anything is finalised. The handwriting on the page is never altered — the doctor remains in control of what's prescribed.

We're still early, and we'll publish real numbers when they're earned and independently meaningful. Until then, the honest version is this: reading Indian clinical handwriting reliably is a genuinely hard problem, and we'd rather get the safety architecture right first.