OCTOBER 2026
•
5 MIN READ
•
Low-Resource NLP
The Boring Solution: How We Anchored Clinical Swahili for Gemma on Budget Androids
Small language models are strong in English and weak in Swahili. When you fine-tune a 2B parameter model on low-resource clinical datasets, it tends to hallucinate plausible-sounding medical terminology that is dangerously wrong. In collaboration with Nama Labs, we built Sema: a morphological anchoring pipeline that strips Swahili prefixes and suffixes before indexing. In document retrieval across the Tanzania Standard Treatment Guidelines, this simple, unglamorous fix elevated accuracy from 54% to 97.4%.
OCTOBER 2026
•
4 MIN READ
•
Edge AI & Architecture
Why We Designed for 'Offline Capable' Rather Than 'Cloud Native' in East Africa
A dispensary doctor at 2am in rural Morogoro does not have 3 seconds to wait for a WebSocket round-trip to an AWS region in Europe. If a digital health tool goes dark when the cellular tower loses power, it isn't an assistant—it's an obstruction. We break down why Nia ships with quantized local model weights (Gemma 3 270M and ggml Whisper), SQLite-based encrypted local storage, and batched asynchronous telemetry that only syncs when connectivity returns.
SEPTEMBER 2026
•
6 MIN READ
•
Clinical Safety
Zero Patient Names: Designing Privacy & Epistemological Safety into CDS
When building machine learning systems around human health, the first design failure is collecting unnecessary identifying data. Nia enforces an architectural invariant: no patient names, no phone numbers, no national IDs can be recorded in our telemetry tables. The system logs condition codes, age brackets, and whether a clinician agreed or overruled an STG suggestion. This empowers epidemiological insight without compromising the sanctuary of the doctor-patient encounter.