Skip to main content

Applied AI Africa · A griot's tale of machines

A3-SERE

Systems, Edge & Real-World Engineering

Every great story begins with a constraint — ours has four. Here you learn to build AI that survives a $30 phone, a signal that comes and goes, four languages in one sentence, and a budget that barely whispers. Not a course of slides. A journey through the fire, proven by what you ship.

Target device
$30–80 Android phone
Network
2G/3G, often none
Languages
Swahili · Luganda · Yoruba · Pidgin
Budget
$5–15 / month, all in

The four non-negotiable pillars

Before a single line is written, the griot states the rules. Every decision in this program is judged against these four constraints.

Hardware Reality
Models run on $30–80 phones and $5–15/month VPS nodes.
Connectivity Reality
Offline-first, sync-on-connect, graceful under 2G/3G.
Language Reality
Swahili, Luganda, Yoruba, Pidgin, and code-switching are requirements.
Cost Reality
No proprietary LLM APIs for core inference; measure cost per 1,000 queries.

Most courses track what you’ve read.
This one works out what you know.

The Compass is a learner model over a 37-skill prerequisite graph. You shouldn’t take that on faith — the whole engine runs in your browser, so use it right now. No account. Answer one question and watch it revise skills it never asked you about.

Live · runs on your device

Three questions. Watch it reason.

This is the real engine on the real item bank — 88 calibrated items across 37 skills. No account, no server, nothing sent anywhere. Answer one question and watch it change its mind about skills it never asked you about.

Why it can be trusted with that answer

An assessment that grades you and won’t explain itself is asking for trust it hasn’t earned. Four commitments, each one checkable.

It diagnoses causes, not symptoms

Your retrieval work is failing. Most systems tell you to re-read the retrieval chapter. The Compass walks the prerequisite graph, finds that you never demonstrated embeddings, and points there instead — naming every downstream skill that unblocks when you fix it.

It refuses to flatter you

Finishing a page is capped at 58% mastery, and a skill it has never measured is capped lower still. Reading is not knowing. Nothing on the Compass goes up because you scrolled to the bottom of something.

It shows its work

Every number opens into the evidence that produced it — prior, each answer, decay since, the inference used. Download the whole learner model as JSON. Disagree with it if you can find the step where it went wrong.

It obeys its own four pillars

The entire engine is arithmetic that runs in your browser. No model download, no API, no tokens — it works offline on a $40 phone at $0.00 per 1,000 recommendations. A tool that taught these constraints while ignoring them would be refuting itself.

The zero-to-master path

The certificate is earned only when you ship a working, publicly reachable system — a rite of passage, not a completion badge.

1
Foundation
Weeks −1 to 0
Primer → Part 0 → self-test. Prove you can build before you begin.
2
Core competency
Weeks 1–8
Eight labs, each gated by a success bar — not by "finished the reading".
3
Ship
Weeks 9–10
One live, benchmarked system that satisfies the four pillars.

What makes this tale different

Most platforms optimize for the notebook. A3-SERE optimizes for the field.

Typical platformA3-SERE
Clean English datasetsCorrupted, multi-currency, code-switched reality
Cloud GPU assumedCPU-first, $30 phone assumed
Accuracy in a notebookLatency, memory, and cost on a live system
Copy-paste that may not runEvery example verified
English-onlyMultilingual by law

Measure what matters

Success is not a notebook accuracy number. It is a live system a real user can reach on WhatsApp or USSD, that runs within the memory budget of a mid-range phone, and remains useful when the network drops. That is how the tale ends — and where yours begins.