Used by learners at
These are open questions. Nobody has answered them at scale. But we already know one thing that used to be a boundary: Benjamin Bloom showed in 1984 that mastery-based instruction lifts the average student one standard deviation — past the 84th percentile of a conventional classroom. Add a one-to-one tutor and the same student rises two standard deviations — into the top 2%. That has been the verified ceiling of human learning for forty years, and it has been unreachable at scale because a private tutor per child is what it takes.
Until now. The mastery-gating algorithm carries the first standard deviation. Carefully-implemented AI — Socratic dialogue, layered explanations, real-time scaffolding — adds another half. An expert human tutor delivers the last half. Three layers, one loop. Not one of the three was possible a decade ago; all three are now. Which means the questions above are, for the first time, open to serious experiment.
Come adventure with us. All content is free forever. You pay only when the AI runs. Bring your own effort; we bring the leading edge of learning theory and the tools to apply it.
Three open questions we are running an experiment to answer
For decades genius has been treated as innate — a fixed trait of the few. But every capacity we've measured so far turns out to be trainable, given the right sequence and the right feedback. If a learner is never blocked, never made to repeat, and never left to guess what to work on next — how far does the capacity extend? We're going to find out.
The nominal 20-year path is set by the pace of the median in a class of forty. Strip out the median-drag, gate advancement on evidence of mastery, unblock hard nodes with a human tutor, and the ladder should collapse dramatically. Nobody knows how far. Our working hypothesis is six years from age 8. We publish what actually happens.
The Fields Medal goes to mathematicians under 40 because the most creative decade of a life is the one before it. If a learner reaches the frontier of their field at 14 instead of 24, they get ten more years of original work. That is the prize. Not a shorter education — a longer creative life.
You don't need to believe our answer. You need to be willing to run the experiment on your own learning, on your own child's learning, and see what shows up. The diagnostic is free. The content is free. The evidence — good and bad — will be published as it comes in. If it doesn't work for you, cancel; the mastery map you built is yours to keep. If it does, tell us what we should try next.
Why
Form 3 concepts your child moves past
A classroom of 40 covers the syllabus at the median. Nobody names the specific six your child never actually mastered. That's where the marks are lost.
what most apps tell you
Not which concepts. Not which prerequisites. Not what to do next. Content volume dressed up as progress. Bloom's ceiling isn't reached by more content — it's reached by knowing what to work on next.
concepts already wired into the graph
Every one is free to read once you sign in. What you pay for is the AI that routes your specific learner through the graph — never for the graph itself.
Method
In 1984 Benjamin Bloom found that one-on-one mastery tutoring lifts students two standard deviations above conventional classrooms. Forty years later this remains the verified ceiling of what learning can produce, and it has been the unreached target of every serious edtech investment. The reason nobody has closed it at scale is that the sigmas come from different mechanisms. ALE ships all three — because none of them alone reaches the ceiling, and every product that pretended one did has empirically fallen short.
Layer 1 · +1σ
→ 84th percentile
A curriculum knowledge graph. Diagnostic that names the specific gaps. Mastery gating that refuses to advance until evidence is in. FSRS spaced repetition. IRT question calibration. This is what a strong tutor does in their first twenty minutes, running continuously in software.
This lifts the average student to the 84th percentile of a conventional classroom. What ALE delivers today.
Layer 2 · +½σ
→ 93rd percentile
Layered explanations generated for the specific misconception you fell into. Socratic chat that walks you toward the answer instead of narrating it. Oral assessment. Cross-domain analogies. This is the piece frontier LLMs make possible for the first time — done carefully, with domain grounding and the graph in the loop, not a chatbot hallucinating vibes.
Adds another half standard deviation — into the top 7%. Now shipping across every subject.
Layer 3 · +½σ
→ top 2%
The last mile — presence, real-time reading of confusion, Vygotskian ZPD calibration, motivation on days learning is hard. Only dispatched when the loop identifies a genuine block the algorithm and the AI together could not clear. A fraction of tutor hours per learner, targeted at the exact node that needs a human.
Closes the final half sigma — into the top 2%. ALE Assist, rolling out via the Talente marketplace.
Everyone who has tried to close 2σ with one layer has failed. Cognitive Tutor and ALEKS delivered the first half of Layer 1 and got ~½σ. Byju's and uLesson dropped the knowledge graph and got less. Khan Academy is retrofitting Layer 2 onto a content-first product. The AI-only edtechs are betting that Layer 2 alone reaches the ceiling — the research does not support that. All three layers, in one loop, is what nobody has shipped before.
How it works
A 45-minute subject-specific diagnostic identifies your current mastery on every concept in the graph. Adaptive branching — a correct answer at level N triggers a level N+1 follow-up; a wrong answer triggers a prerequisite probe.
Your mastery map is generated. Every concept is scored — mastered, partial, gap — and root-cause gaps are traced backward through the DAG. The frontier — what you're ready to work on next — becomes your personalised sequence.
You work through the sequence. Problems scaffold from prerequisite to target, with hints declining as mastery evidence accumulates. FSRS spaced-repetition maintains what you've closed. The engine never asks you to review what you already know.
Weekly the diagnostic runs on the worked nodes. The map updates. Movement is visible on the parent dashboard. If a node stays blocked after three re-tests, ALE Assist packages the stuck point for a human tutor.
The K-12 → PhD path is nominally 20 years because a classroom has to move at the median, repeats what most learners already know, and cannot name — let alone close — the specific gap a specific child is carrying. Strip out the median-drag and the repetition, gate advancement on mastery evidence, unblock the hard nodes with a human tutor, and the same ladder should collapse. How far? Nobody has run this experiment at scale before. The path below is the ALE architecture's outer promise — the ceiling we're actively working to reach, not a guarantee we've delivered. What a child actually achieves depends on effort, hours, and support, and we'll publish the evidence as it comes in.
Year 1
Age 8 → 9
Diagnostic locates the child's real position (often 1–2 grades below nominal). Adaptive engine closes the numeracy, literacy and Kiswahili L1 gaps most classrooms leave behind.
Year 2
Age 9 → 10
Full CBC junior-secondary science, math, humanities. The graph is now dense enough that root-cause remediation is the norm, not the exception.
Year 3
Age 10 → 11
Formal KCSE mathematics, physics, chemistry, biology begin. Kiswahili L2 pathway available for English-first learners. FSRS scheduling begins across the entire mastered graph.
Year 4
Age 11 → 12
The full KCSE syllabus completes. Mock papers, exam-date-anchored study plan, viva-style oral assessments. Grade-A ceiling is available to any learner who has ridden the loop honestly.
Year 5
Age 12 → 13
First-year foundation content bridges into the discipline of choice — Da Vinci Engineering, MBChB Years 1–2, CS, econometrics. Hybrid engine begins: DAG core + early Semantic Concept Graph.
Year 6
Age 13 → 14
ALE Advanced takes over: Semantic Concept Graph with Heuristic Mapping, methodological fluency diagnostics, literature-map navigation, viva preparation. The learner is now doing original research, not consuming curriculum.
The four compressions that make six years possible
1 · No wasted time
FSRS never asks the learner to review what they already know. The classroom, by necessity, does.
2 · No hidden gaps
Root-cause tracing names the specific unclosed prerequisite behind every failure — and closes it before advancing.
3 · No blocked days
When the algorithm can't unstick a node in three sessions, ALE Assist puts a real human tutor on it within the hour.
4 · Right ontology per rung
DAG for convergent K-12, Semantic Concept Graph for rhizomatic postgrad. Neither gets forced to pretend it's the other.
Why this isn't rote learning
The crammer apps optimise for one thing: getting the specific past-paper question right next time. That's memorisation with a leaderboard. Understanding is different — it transfers to a question the learner has never seen. Every mechanism in ALE is chosen to build the transferable version.
Questions are IRT-calibrated to your current ability — always one step past what you've already got. If you can only answer questions you've memorised, the engine bumps difficulty until you have to actually reason. There is no leaderboard for grinding the same item.
A wrong answer surfaces a Hook / Rule / Bridge / Trap explanation — the intuition, the underlying principle, the connection to what you already know, and the specific misconception you fell into. Not just "the answer is B."
The tutor chat asks questions back. If you're stuck, it walks you toward the answer by making you notice things — the way a strong human tutor would — not by narrating the solution while you nod.
Fail a Form 3 target concept and ALE walks the graph backward to find the actual root gap — often two years upstream. Drilling harder on the target never fixes the upstream cause. The whole point of the DAG is to make that hidden dependency visible.
The Semantic Concept Graph deliberately does not gate. It illuminates the neighbourhood of methods, literature and cross-disciplinary connections the researcher needs to weave together. Doctoral work is not "master this list" — and we don't pretend otherwise.
FSRS spaced repetition schedules review just before you're about to forget — measured per concept per learner. If a concept doesn't stick past exam day, it wasn't understood. The whole graph is designed to compound over years, not decay after a term.
What's been tried before us
We are not claiming to have invented adaptive learning. It is a mature field with a real track record — and a real ceiling that has held for decades. What is new is the combination the last few years have made possible: knowledge-graph density at the scale a full K→PhD curriculum needs, spaced-repetition schedulers that are genuinely per-learner (FSRS, shipped 2022), IRT that runs online, and frontier LLMs that can generate layered explanations and Socratic follow-ups on demand. Those pieces did not exist together before. That is what makes this a fresh experiment worth running.
| Programme | Era | What it tried | Effect / outcome |
|---|---|---|---|
| Bloom's original studies | 1984 | 1:1 human mastery tutoring vs classroom | The 2σ result — the ceiling everyone since has been chasing |
| Carnegie Learning · Cognitive Tutor | 1990s → today | Model-tracing intelligent tutor for algebra, Pittsburgh schools | ~0.3–0.4σ gains in RCTs; strongest peer-reviewed adaptive-learning evidence at scale |
| ALEKS | 1990s → today | Bayesian knowledge-state assessment, university math placement | Widely deployed in US higher-ed; gains modest but reproducible |
| DreamBox Math | 2006 → today | K-8 adaptive math with strategy diagnosis | Positive but small effects in independent studies; content-limited |
| Knewton | 2011 → 2019 (sold, wound down) | White-label adaptive engine for university textbooks | Overpromised on personalisation; commercial failure; instructive on what "AI-powered adaptive" cannot substitute for structure |
| IBM Watson Classroom | 2016 → 2019 (killed) | AI teacher-assistant for K-12 | Discontinued after failing to demonstrate outcomes |
| Byju's | 2015 → 2024 (collapse) | Video-first adaptive app, India-scale | Content-first architecture without a real mastery graph; unit economics + outcomes both failed |
| Squirrel AI (China) | 2014 → today | Fine-grained knowledge-point mastery, 1:1 chain schools | Vendor-reported gains; independent replication scarce |
| Khan Academy · Khanmigo | 2023 → today | Content library + GPT-4 Socratic tutor | Content-first, adding conversational LLM; not knowledge-graph-first |
| Duolingo | 2012 → today | Well-tuned SRS + gamification, single domain | Strong retention mechanics; deliberately narrow scope |
The pattern is honest: adaptive engines built without a full-ladder knowledge graph plateau around ½σ. Engines with a graph but without human unblock hit the same ceiling for a different reason. Nobody has combined all three — dense graph, calibrated adaptive loop, real-human unblock on the residual — at population scale, in a curriculum aligned to a specific country's exam boards, in a language other than English. That is what we are testing.
How we differ
| Category | Example | What it optimises for | How ALE differs |
|---|---|---|---|
| Content library | Khan Academy · Coursera · edX | Availability of good videos | Graph-first not video-first; every concept is a node in a mastery-gated DAG aligned to CBC / KCSE / IGCSE; layered explanations generated per-learner, not the same video for everyone |
| Language app | Duolingo · Pimsleur · Fluently AI | Vocabulary retention (Duolingo) · oral fluency via graduated audio drill (Pimsleur) · adaptive AI conversation with pronunciation + grammar feedback (Fluently, English only) | ALE incorporates all three mechanics — SRS + streaks (Duolingo), graduated audio anticipation drill (Pimsleur), real-time AI conversation with sound-by-sound pronunciation and grammar correction (Fluently) — and adds what none of them ship: KCSE-competent Kiswahili as L1 AND as a zero-to-hero L2 pathway for English speakers, full grammar (ngeli, tense/aspect, composition), the Kamusi dictionary integrated as a one-tap popover, a scientific-Kiswahili corpus for STEM in Kiswahili, and cross-curricular integration so Kiswahili sits inside the same learner's mastery record as their other 10 subjects |
| Video-first adaptive | Byju's · uLesson | Consumption metrics dressed as personalisation | Real IRT + FSRS running per learner, per concept; wrong answers produce Socratic follow-ups not more videos; mastery gating not passive viewing |
| Exam crammer | Zeraki · Elimu · PastPapers.co.ke | Past-paper coverage | Root-cause tracing across the DAG — a Form 3 fail traces back to the specific Form 1 gap and fixes it first; understanding transfers, past-paper drilling doesn't |
| LMS | Google Classroom · Moodle · Kenya Ed Cloud | Delivering content the teacher assigned | Not an LMS at all — plugs into one when the school has one; the adaptive loop runs per learner regardless of what the class is doing |
| Private 1:1 tutor | The historical way to close 2σ | Presence, per-child diagnosis, real Socratic dialogue | Algorithm carries the mastery load across every subject at once; human tutor is dispatched only when the loop identifies a genuine block — a fraction of tutor-hours, available on demand |
| Adaptive engine, English-only | ALEKS · Cognitive Tutor · DreamBox | US / European curricula | African curriculum built-in (CBC, KCSE, IGCSE-Kenyan-context); Kiswahili STEM authored not translated; MBChB / USMLE Step 1 for East African medical schools |
The moat is not one piece. It is the combination: dense knowledge graph + IRT + FSRS + layered LLM explanations + Socratic tutor + human unblock + African-curriculum alignment + Kiswahili STEM. Every competitor has some of the parts. Nobody has all of them wired together per learner.
K → PhD, with the right ontology per rung
Convergent knowledge — arithmetic → algebra → calculus — has genuine prerequisites, and a mastery-gated Directed Acyclic Graph is the correct instrument. Postgraduate research is rhizomatic, connected by relational edges (is-methodology-for, theoretically-contradicts, serves-as-literature-for). A DAG gets the first right and forces the second into a shape it doesn't have.
| Rung | Knowledge structure | Engine | Adaptive mechanism |
|---|---|---|---|
| Class 1–6 (CBC) | Convergent, cumulative | DAG mastery loops | Gate on mastery evidence |
| Grade 7–9 (JS) | Convergent + subject depth | DAG mastery loops | Per-subject mastery path |
| Form 1–4 (KCSE / IGCSE) | Convergent, exam-scoped | DAG mastery loops | Prerequisite cascade + FSRS |
| Foundation undergrad | Bridging + mirror of secondary | DAG mastery loops | Root-cause remediation |
| Advanced undergrad | Convergent core + lateral | Hybrid DAG + Semantic Graph | Core gates + lateral suggestions |
| Postgraduate / PhD | Rhizomatic, relational | Semantic Concept Graph | Heuristic mapping, no gating |
A postgraduate doesn't need to "master chapter 7 before chapter 8." They need to see how instrumental variables relate to their identification strategy, which literature their design extends, and where their methodology sits in the neighbourhood of adjacent work. The Semantic Concept Graph illuminates. It doesn't gate.
Curriculum
CBC and KCSE for Kenyan learners, IGCSE for the international/private-school cohort, foundation-year undergraduate content for Kenyan universities. Kiswahili is first-class — not a translation layer, an authored curriculum for L1 speakers and a zero-to-hero L2 pathway for English speakers.
Grade 1 → KCSE → undergraduate → postgraduate mathematics
CBC → KCSE → undergraduate → postgraduate physics
CBC → KCSE Papers 1–3 → MBChB biochemistry
CBC → KCSE → MBChB physiology & Step 1
Physical, human, and economic geography
African, Kenyan, and world history
MBChB Years 1–3 + USMLE Step 1 clinical vignettes
Foundation → Da Vinci Engineering undergrad
Algorithms, data structures, first-year CS
L1 mother-tongue + L2 pathway for English speakers
CBC → KCSE literature & language
Who it's for
An 8-year-old can pick a username and PIN and start. A grown-up gets an email — no account required for children too young to have one.
Kids sign up →Every KCSE subject with mocks, mastery tracking, and an FSRS scheduler that plans backwards from your exam date. IGCSE core and extended syllabi supported.
Sign in →Foundation-year remediation, clinical vignettes, engineering, CS. Depth without noise. Layered explanations for every concept.
Get started →ALE Advanced runs on a Semantic Concept Graph with Heuristic Mapping — literature, methodology, and cross-domain synthesis. Not a mastery ladder.
Preview →ALE Assist · rolling out
Every adaptive app claims "AI-powered personalisation." Parents have been burned by hollow AI claims. ALE Assist is the anti-cargo-cult signal: when the adaptive loop detects a learner stuck on a specific concept — three spaced sessions with full hints, no advance — it packages the stuck point as a scoped 45-minute micro-task and dispatches it to a vetted human tutor via Talente. The session runs in an embedded whiteboard; on completion, ALE resumes the loop from the unblocked node.
A fraction of learner time. A fraction of tutor-hours. This is what makes population- scale approach to Bloom's second sigma economically possible.
Contract
How we charge · why content is free
Every lesson, worked example, past paper, and Kamusi entry is authored once and hosted at near-zero marginal cost. Every diagnostic run, every adaptive question the model selects, every layered explanation and every tutor-chat turn burns real inference time on real GPUs — a cost per invocation that we cannot fake away. So we don't. Content is free. AI is metered, priced at what it actually costs us plus a modest routing margin.
Free with a signed-in account
Metered — pay per invocation
Why we require a signed-in account for content
The graph, the lessons, and the worked solutions cost real effort to author. Free content behind a sign-in is not paywall friction — it's the only way to keep the whole library free for actual learners while stopping competitors and scraper farms from siphoning years of authoring work in a weekend. Every account is rate-limited per session, every request is bound to a learner-ID, and there is no bulk-export endpoint. Learners can export their own notes and mastery record any time; nobody gets to export ours.
Plans
Unused credits roll over. Bolt-on credit packs available any time. If your first paid month doesn't move your mastery map on the re-test, we refund the month.
KES 0
signed-in account
Unlimited content. 30 AI credits/month — enough for a full diagnostic in one subject.
Create free accountKES 800/mo
single subject focus
Unlimited content. 300 AI credits/month for one subject's adaptive drilling.
Choose LiteRecommended
KES 1,500/mo
all subjects
Unlimited content. 800 AI credits/month + 2 ALE Assist human-tutor credits.
Choose LearnerKES 3,500/mo
up to 4 learners
Unlimited content. 2,500 pooled AI credits + 8 ALE Assist credits.
Choose FamilyCredit top-ups: KES 200 for 100 credits; KES 400 for one ALE Assist human-tutor session (45 min). No auto-billing, no surprise charges — the app shows the credit cost of every action before it runs. Institutions: per-learner metered plans with pooled credits and admin dashboards. Talk to us.
Forty-five minutes. A readable mastery map: per-topic scored, per-prerequisite gap flagged, per-concept next step named. More useful than most schools' end-of-term reports. Free even if you never subscribe.
The diagnostic re-runs after your first paid month. If it doesn't show measurable movement on the mastery map, the month is refunded. The risk is on us, not on you.
Every concept you master and every gap you close is yours. Full export at any time. GDPR-aligned. Kenya Data Protection Act compliant. Never sold, never used for cross-platform profiling.
Start where you actually are
Free diagnostic. No credit card. No app to install. Cancel any time — the mastery map is yours to keep.
The Adaptive Learning Engine
In 45 minutes we'll map exactly what your child knows — every mastered concept, every specific gap, every prerequisite hidden two grades upstream.
The map is free. Yours to keep. Then decide what to do with it.
No card. Kids can sign up on their own. Cancel any time — the map stays yours.
Preview
Every screen ALE ships is anchored on the same thesis: the specific concept, the specific gap, the specific next step. No "47 lessons completed" abstractions.
Screen 01
Every concept in the subject DAG, scored. Green nodes are mastered and being maintained by spaced repetition. Amber nodes are partially mastered. Red nodes are identified gaps — with the specific prerequisite that's blocking them named.
A parent's read: "Chemistry Form 3 — Aisha has mastered 18 of 28 concepts. Her three red gaps are all downstream of the mole-concept prerequisite from Form 2. That's the node to close first."
Aisha · Form 3 · Chemistry
Mastery map — 18 of 28 concepts
Root-cause trace · Stoichiometry (failed)
Stoichiometry — target
Failed 3 items · IRT θ = -1.2
walks the DAG backward
Mole ratios
Partial — 40% correct
still upstream…
The mole concept — root cause
Drill sequence begins here
Screen 02
When you fail a target concept, ALE doesn't drill you harder on the target — it walks the DAG backward and finds the root-cause node. Drilling a Form 3 topic can't fix a Form 1 gap; the engine locates the Form 1 gap and closes it first.
This is where 1σ actually comes from. Not more content — better routing.
Screen 03
Item Response Theory calibrates every question to your current ability estimate. If you're strong you get harder items; if you slip you get scaffolded ones. Wrong answers trigger a layered explanation — Hook, Rule, Bridge, Trap — not a red X.
Behind the surface: FSRS predicts when you're about to forget each concept and schedules review just before that moment. You never review what you already know.
A 0.500 mol sample of calcium carbonate is heated until it fully decomposes. What mass of calcium oxide is produced?
Layered explanation · Bridge
One mole of CaCO₃ produces one mole of CaO. So 0.500 mol of CaCO₃ yields 0.500 mol of CaO. Molar mass of CaO = 40 + 16 = 56 g/mol. Mass = 0.500 × 56 = 28.0 g. The classic trap on this item is to reach for the molar mass alone (56 g) — that's mass of one whole mole, not of the half-mole you have.
Kid mode · Class 4
Kito says: nice work!
You closed 3 concepts this week 🌱
Adult mode · KCSE Form 4
Study plan → KCSE 2026-10-15
FSRS-projected — 12 concepts drop below threshold by exam
Screen 04
A Class 4 learner sees a lion cub called Kito, oversized buttons, sound cues, XP, streaks, and a Memory Garden that grows as concepts close. A KCSE candidate sees layered explanations, IRT ability curves, an FSRS study plan calibrated to the exam date, and a portfolio the university can read.
Same underlying concepts. Same mastery record. The mode switches; the map compounds over years.