ShamwariAI

Blueprint

How Shamwari is built

Shamwari isn't one model behind an API — it's three, and which one answers a question depends on whose question it is.

Cloud

Runs on routed third-party models. Fast, general knowledge, no persistent memory of you. Handles anything that isn't about your own life.

Ground

The Zimbabwean corpus — law, tax, regulation — kept fresh and cited by section. What makes an answer correct instead of just fluent.

Mind

A small, open-weight model that runs on your own device. The only thing that ever sees your personal data, because it never has to leave.

Data sovereignty

The two rules that must not be broken

A question arrives

A general question
A question about you
flows into

The gate — checked twice

Parse request

reads the scope the request claims — personal, community or platform

Edge checkgateway/src/scope.ts

fast fail, saves a Core round trip

Core checkcore/main.py::resolve_scope

authoritative — the edge can't override it

Route decision

Cloud, or refuse and point at Mind — never a silent downgrade

routes to

Where it goes

Shamwari Cloud → routed provider
Shamwari Mind → on your device

> POST /v1/chat/completions {scope: "personal"}

409 scope_requires_local_inference

> POST /v1/chat/completions {scope: "community"}

200 → routed to Shamwari Cloud (Qwen)

A question about your own life is answered on your device by Shamwari Mind. It is checked twice — once at the edge, once in the service that owns the data — because a leak here isn't an outage, it's a broken promise.

1Your personal data never reaches a third-party model

Not the cloud providers we route to, not anyone. If a question is about your own life and the only available answer would mean sending it to an outside model, Shamwari says so and stops — it does not quietly answer with a shrug instead.

2Only open-weight output trains Shamwari Mind

The model we train and give away is never built from output whose licence forbids that. Every exchange is labelled the moment it's generated, and the training pipeline physically cannot ingest anything labelled otherwise.

Read the full technical detail

Ground — the moat

Freshness beats volume

A grounded small model beats an ungrounded large one on "what is the current PAYE threshold" — because the large one is guessing from memory and the small one is reading. Ground is a living corpus with an effective date on every document, not a one-time scrape: when a Statutory Instrument supersedes an older one, retrieval stops citing the repealed version.

What people actually ask becomes the roadmap for what we ingest next —shamwari.groundMisseslogs every question Ground couldn't answer, so the corpus roadmap is written by the people asking, not by guesswork.

  1. Law — the Constitution (approved, first in), ZimLII, Veritas, Parliament
  2. Economy — ZIMRA, RBZ, ZIMSTAT, Ministry of Finance
  3. Verticals — AGRITEX, MoHCC, ZIMSEC (licence required)
  4. Language — Masakhane, Common Voice, VOA, UZ ALRI partnership

Five sources are blocked pending licence review — nothing gets scraped first and asked about later. Seethe open issuetracking that.

Mind

How the model we give away gets trained

Distillation, not fine-tuning from scratch — and the one step in this pipeline that's the actual asset.

  1. 1

    Seed question/answer pairs from Ground documents.

  2. 2

    Generate candidates in Shona, Ndebele, Zimbabwean English and code-switched registers with Kimi K3 and Qwen — never Claude or GPT, whose terms bar using their output to train a competing model.

  3. 3

    Correct with paid human annotators. This is the irreplaceable part: Shamwari employs Zimbabweans to teach an AI Zimbabwe.

  4. 4

    Train with QLoRA on Qwen 4B — an Apache-2.0 base, so Mind is genuinely open: weights, data recipe and eval suite all publishable.

  5. 5

    Quantize to q4, targeting a ~US$120 Android device.

ShamwariBench — a Zimbabwean law, tax, agriculture, health, Shona and Ndebele fluency benchmark — gets built and published beforetraining starts. Whoever owns the benchmark owns the argument about who's actually best at African languages.

Track Mind on GitHub

Cloud economics

Where the margin actually comes from

AI Gateway passes provider pricing through at cost — there's no margin in simply reselling access, since anyone could go direct and undercut it. Margin gets built from three places instead:

  • Route the bulk to cheap models. Qwen and DeepSeek handle most traffic; Kimi K3 for reasoning and long context; premium models are opt-in only. Target: at least 70% of tokens on the economy tier.
  • Cache hard. Exact-match caching is free on AI Gateway, and in a market where thousands of people ask near-identical questions about ZIMRA thresholds, that hit rate should be high.
  • Let Ground do the lifting. A grounded 4B answer beats an ungrounded 400B answer on a question with a factual, checkable answer. Correct answers from cheap models is the whole game.

Ecosystem

Where Shamwari sits in the Bundu Ecosystem

Bundu Foundation is a Zimbabwe-registered non-profit that holds the Shamwari IP and governs the wider ecosystem. Shamwari is sold commercially under Nyuchi Africa; the code itself stays Apache-2.0.

Bundu Labs

Research.

Mukoko

The consumer super-app — several sub-apps currently running in beta.

Nyuchi Africa

The commercial pillar. Shamwari's paid tiers run here.

Shamwari AI

The community pillar — this site.