The platform
A WhatsApp-based, AI-powered upskilling platform
Our WhatsApp Learning System (WLS) blends local social infrastructure with mindset work and project-based learning. The design decision underneath it is simple. In a December 2025 survey of Bamako youth, 99.1% used WhatsApp on their own phones, 88.5% spent more than 30 minutes a day on it, and 71% named it their main source of job and program information (Bamako survey, n=336).
Over six months in 2025, learners sent the AI Mentor about 23 times more messages on WhatsApp than in the app (6,618 vs 284). The app stays, as the depth channel where richer work happens; it holds a 4.7/5 rating on Google Play from 1,400+ reviews (November 2025). WhatsApp is the scale channel, and 75–80% of learners who start there later transition into the app.
Bamako survey, n=336 | December 2025
Six-month platform data | December 2025
1,400+ reviews | Google Play | November 2025
AI Mentor
An instructional facilitator, not a chatbot
Kabakoo has been building AI for its learners since 2019 and has operated an AI coach on WhatsApp since 2023. The AI Mentor doesn't wait for questions about course content. It reviews actual learner work, a project video or a business plan, and returns personalized feedback, in French and Bamanankan text, 24 hours a day.
An analysis of 11,171 bilingual conversations (June 2026) shows what a mentor in two languages actually does. Most learners who used Bambara also used French, switching by task: digital and technical questions came about four times more often in French, while entrepreneurship and endogenous knowledge ran twice as often in Bambara. Language doesn't just provide access. It changes what becomes thinkable.
When Kabakoo first deployed LLM-generated personalized feedback over WhatsApp in November 2023, monthly learner video submissions went from 667 to 870 within the month. Feedback that arrives, in your language, at 11pm, changes whether work gets made at all.
Nearly 40% of voice interactions in Bamanankan | June 2026
In one month | Platform data | November 2023
Peer groups
Grins, because tea circles already work
A grin is the Sahelian tea circle, the place where young people already meet and back each other. Kabakoo's in-person peer groups carried the generic name "Clubs" until January 2026, when they were reframed as grins in a four-week experiment with the 1,000 most active learners. In-person attendance went from 29.5% to 70.5%.
While this is a descriptive before/after comparison, it still taught us the lesson we keep relearning. Cultural anchoring is more than branding. Inside the WhatsApp Learning System, the same social layer is productized as pods, described below.
In 4 weeks | Descriptive before/after, N=254 | January 2026
Entry phase
One doable task a day, for 30 days
The 30-day challenge is the entry phase of Kabakoo's current cohort program. One short, actionable task arrives each day over WhatsApp; tasks unlock daily, so the cohort moves roughly together.
Pods, AI-matched peer groups
Pods bring learners into small groups of three to five, matched on shared interests and geographic proximity, to work on concrete projects.
In-product economy
Kabakooins
Kabakooins are the platform's tokens, earned through learning activity and spendable on things like equipment rental or peer support. They're live in the product.
African languages
Starting with Bamanankan
Among surveyed Kabakoo learners, 92% have difficulty reading Bamanankan and half can't read it at all, while learners consistently prefer oral Bamanankan. In the AI Mentor's conversation data, text stays overwhelmingly French while nearly 40% of voice interactions happen in Bambara. Text-based AI in Bamanankan, which the AI Mentor already does, only covers part of the need. Voice is the real frontier.
Here's where that frontier actually is. The best word-error rates in the Bamanankan speech-recognition ecosystem sit around 46–48% on clean audio. Our own Whisper fine-tune moved WER from 0.98 to 0.64. We are building the raw material the whole field lacks: a dataset of clean single-voice Bamanankan audio and clean French–Bamanankan sentence pairs, plus a learner community that generates real code-switched speech every day.
Learner surveys | 2025
Obviously not yet production-ready
Clean audio, sentence pairs
Highdigenous pedagogy
Mindset first, then skills, then real projects
While technical skills like web design are optional, mindset work is mandatory for all learners. Cohorts run 16 weeks: mindset modules in weeks 1–6, then technical skills applied through KLIPs, Kabakoo Local Innovation Projects, then a final public showing of learner work. Over 200 hours per learner.
The measurement framework was designed with Prof. Dr. Martin Schneider's team at the University of Paderborn; the mindset modules were built in conversation with Dr. Catherine Thomas (University of Michigan) and Dr. Richard Sedlmayr. Findings on growth mindsets and employability in Mali were published in the Academy of Management Proceedings (2023). A second study on Highdigenous pedagogy is under review at the International Journal of Educational Development, with a public preprint on SocArxiv.
Evaluating the AI
Four levels, four kinds of evidence
Kabakoo assesses its AI-enabled system against the four-level framework built by the Agency Fund with the Center for Global Development.
Did the AI do what we intended?
Automated evals score the mentor's outputs for empathy, groundedness, concision, language quality, cultural relevance, safety, and escalation. Non-engineers write test cases too. One payoff: a model swap matched accuracy at 57% lower inference cost.
Could people actually use it?
A daily dashboard tracks acquisition, activation, engagement, and learning progress, with automated data-quality tests.
Did something change in the learner?
Validated surveys on self-efficacy, grit, growth mindset, and social standing, fielded with each cohort, plus portfolio and reflection analysis.
Did lives improve, and what caused it?
A longitudinal panel with recruit-and-delay controls measures income, savings, and social capital over three years.
The framework is not an escalator. A model can pass while the product fails; a product can retain attention without producing learning; a learner can gain confidence without durable economic improvement. Each transition gets tested. The essay "The AI worked. Did it work for the learner?" walks through all four levels in detail.