Research

Sema

Structured bilingual clinical NLP

The problem

Modern language models are overwhelmingly trained on English-language data. When you need a model that can reason about clinical scenarios in Swahili, you have two standard options: fine-tune a large model on Swahili clinical data (expensive, data-scarce, and too large to run on a phone), or translate Swahili input to English, process it, and translate back (lossy, slow, and introduces translation errors that are invisible to the end user and dangerous in a clinical context).

Neither option works when your deployment target is a budget Android phone with no connectivity in a dispensary that sees 80 patients a day.

The approach

Sema takes a deliberately different path: what we call "The Boring Solution". Instead of asking a generative model to translate clinical Swahili into fluent English and back, which fails silently and dangerously when it's wrong, Sema anchors. It strips Swahili input down word by word to its dictionary root, tags the grammar, attaches a plain gloss, and hands that structural skeleton to a reasoning model. The output returns through controlled bilingual templates rather than hallucinated prose.

We publish our results openly, including negative ones. We found that past certain model sizes, morphological anchoring added little to conversational fluency. However, in document retrieval, stripping Swahili's grammatical prefixes and suffixes before indexing took clinical guideline retrieval benchmarks from barely passing to near-perfect. That finding directly powers Lucy's instant STG lookups on edge devices.

Status

Sema is active, ongoing work. The core pipeline is functional and powers Lucy's bilingual capabilities in the current ClerQ prototype. We continue to expand the structured representation vocabulary and improve anchor accuracy across clinical domains.

Rafiki

Edge-native digital twin training framework

The problem

Most AI systems are trained in datacenters on large, centralized datasets, then deployed to environments that look nothing like the training data. In healthcare, this gap is dangerous: a model trained on hospital data from Nairobi or London will behave unpredictably when deployed in a dispensary in Mtwara with different disease prevalence, different drug availability, and different clinical workflows.

The approach

Rafiki is a digital twin training framework designed for edge-native systems. Instead of training once and deploying everywhere, Rafiki enables models to learn from the specific environment they're deployed in. It creates a digital twin of the local clinical context: the disease patterns, the available resources, and the workflow rhythms, using that twin to continuously refine the system's behaviour without sending patient data anywhere.

All computation stays on-device. All data stays local. The digital twin evolves with the facility, not against it.

Status

Rafiki is in active development. The framework architecture is established and we're working on validation across different facility types. This is early-stage work, we're transparent about that, but the direction is clear and the collaboration with Nama Labs continues to be productive.

Declarative guideline verification and clinical Model Context Protocol

The problem

Clinical AI systems often embed diagnostic and treatment rules directly in complex neural weights or opaque codebases. When national guidelines (like Tanzania's Standard Treatment Guidelines and NEMLIT) change, updating black-box models requires expensive retraining and introduces unpredictable behavioral shifts. Furthermore, external regulators and health authorities cannot inspect how a conclusion was reached line-by-line against statutory protocols.

The approach

Hakiki completely decouples clinical rule logic from application software. Every symptom weight, threshold, and contraindication is encoded as declarative, versioned data mapped verbatim to page and paragraph citations of Tanzania's STG. Hakiki's confidence scoring uses deterministic bar-crossing ratios rather than machine-learned probabilities: clinical severity saturates immediately upon encountering red-flag criteria rather than awaiting majority votes of symptoms.

Additionally, Hakiki exposes a clinical Model Context Protocol (MCP) server. External AI agents (such as Claude, Cursor, or custom IDEs) can query Tanzania's STG rules directly through an auditable protocol layer that proposes candidate updates for human medical expert sign-off, decoupling who can reason over national guidelines from who built the original application.

Status

Hakiki is currently request-only for research partners, regulatory bodies, and clinical collaborators. It serves as the internal validation harness for Lucy's underlying reasoning rules.