Akaalo · Nambi

An AI tutor that runs on the device.

Nambi runs a quantised language model locally, on low-cost Android tablets, with no internet connection. It is built for classrooms where the network cannot be assumed.

A Nambi session running on a device in a Ugandan vocational classroom Nurture Africa VTC · Nansana, Uganda

This page is a field report, not a product page. It describes what Nambi is, how it runs, where it is deployed, and what remains unresolved.

§ 01What Nambi is

Nambi is an offline AI tutor. A student talks to it in ordinary language and it teaches, asks questions and adapts to the answers. The difference from most AI tools is the assumption underneath: Nambi does not expect the internet to be there.

It is used as a supplementary ICT module in vocational training, teaching the practical digital skills employers ask for. The same engine is subject-agnostic; the curriculum is what changes.

§ 02How it runs

The model runs on the device itself. There is no API call, no cloud service and no subscription after setup. A session looks like this:

Device→ Local model→ Curriculum retrieval→ Learner

Because everything is local, the tutor keeps working when the connection drops, when data runs out, and when there is no power to charge a router. That is the environment it was designed for.

§ 03Hardware

Device
Low-cost Android tablets, 3 GB RAM and up
Model
Custom model, fine-tuned from Qwen (2B, 4-bit Q4)
Network
None required after setup
Cost to run
$1.30 per student, per term

The design target is the cheapest hardware that still gives a usable tutor. Keeping the model small and the retrieval local is what makes that possible.

§ 04Offline architecture

The model is a custom build, fine-tuned from an open Qwen base for this setting. Nambi combines it with a local curriculum store and retrieval: the curriculum provides the knowledge, while the model provides reasoning, explanation and dialogue. When a student asks something the model is unsure of, the system falls back to the lesson material rather than inventing an answer.

Progress is written to a local database on the device. Reports are generated without a network request and exported for programme reporting.

§ 05The learning system

Learning is conversational. Students ask questions, get explanations in plain language, answer checkpoints, and move at their own pace. The tutor adapts the level of the explanation to the learner and flags where someone is struggling.

§ 06Classroom deployment

Nambi is deployed at Nurture Africa Vocational Training Centre in Nansana, Uganda, across a pilot of 154 vocational students. Sessions run on shared tablets; no internet connection is used.

Outcome evaluation is in progress. Metric definitions, measurement periods and methodology are documented and shared with funders and research partners on request, and published on the pilot page as they are confirmed.

§ 07Current limitations

§ 08Research questions

  1. How well does a small, local model teach a practical skill compared with a cloud model, once connectivity is removed from the equation?
  2. What can African-language interaction data, collected in the field, contribute to models that are not built from English internet text?
  3. What is the real cost of sustaining offline learning infrastructure over several years, including hardware failure and maintenance?
  4. How does learning change when the tool does not depend on a connection, a subscription, or a vendor staying online?

Work with us.

Nambi is deployed, documented and open to research and funding partnerships. If you want to see the evidence, deploy it, or help build it, get in touch.