Bangladesh's own clinical AI

Four reasons compound. Cost: frontier inference on every consultation cannot support free care for patients at national scale; the mission's economics require near-zero marginal inference for routine traffic. Privacy and law: PDPO 2025 treats health data as a specially protected category with localization requirements, and the honest fix for sensitive voice data is not better scrubbing but not sending the data out at all. Language: frontier models are competent in standard Bangla and unusable in Sylheti, Chittagonian, and Noakhali dialect speech, the languages in which Glyph's actual patients describe their actual symptoms. And dependency: national health infrastructure whose intelligence lives behind another country's terms of service is infrastructure on loan.

KhaM-Med builds on Google's MedGemma family, open-weights clinical models that can legally run on servers in Dhaka, with published performance among the best open medical models at a fraction of frontier cost. The base knows board-exam medicine in English. KhaM-Med's work is the distance from there to Glyph's floor: Bangla and its dialects, the local prescription culture, local brand-to-generic mappings, the attendant-mediated encounter, local disease patterns, local report formats.

The training corpus is what the network rightfully owns and licenses: consented, de-identified production encounters, every doctor-corrected note, every verified imaging read, alongside licensed and open medical literature. Only the network that owns the encounters can ever have that data, and everything the model learns stays in the country.

Clinical safety gates everything. Drafts for verification, the doctor signs every note, escalation stays conservative, and no task moves to KhaM-Med until it matches the frontier baseline on locally built evaluation sets: real Bangladeshi films from real machines, real dialect transcripts, real chamber notes.

The path is staged honestly. Today frontier models carry the complex reasoning while every consented encounter builds the corpus. Next, fine-tuned models take the structured tasks and, most importantly, the sensitive flows move to in-country processing, the milestone the privacy constraints are waiting on. The target state routes the large majority of routine inference through KhaM-Med in Bangladesh, with frontier models retained for the rare and the complex. The name carries the family: KhaM is the initials of Khayer and Mamataj, and the model exists so that what twenty crore people teach it stays theirs.