The AI deep stack
What we're building on the AnTeraju spine next — the frontier ideas, honestly labelled as workshop or concept. None of this is shipping today; all of it is designed to be provable when it does.
Nine ideas, zero promises they can't keep
Each card carries its real state. In the workshop means we are drawing it now; Concept means it is designed on the whiteboard and owed a build.
Step-level error attribution
The camera reads the working, not the answer. A student photographs a page of maths; on-device OCR and a reasoning chain locate the exact line where the thinking broke — not “wrong”, but “the sign flipped at step three, and this same slip is in your record from Form 2.” That difference is the whole distance between a tutor and an answer key.
Speak the lesson back
The oldest test of knowledge is the cheapest to automate and the most expensive to administer by hand: can you explain it? Socratic voice sessions in Bahasa, English, Mandarin or Tamil, where the mind asks follow-up questions until the explanation actually holds. The oral exam was always the real standard — only cost made it rare.
The five-year memory
Today's school system forgets a child at every year boundary. AnTeraju's learner model remembers every probe since Year 1, spaces reviews on a spiral that genuinely comes back around, and warns Form 4 Add Maths that Form 2 algebra is quietly shaky — before the exam exposes it, not after.
Kampung STEM
Vectors through batik geometry, thermodynamics through the palm-oil mill, probability through the durian auction, kinematics through the motorbike shop. The syllabus stays national; the context turns Malaysian. The most underrated motivation engine in education is a child meeting their own life inside the lesson.
Pocket Labs
Chemistry you can hold: an AR molecule playground rendered by a RM90 Pi in the school's own cabinet, a drone simulator that flies with no internet, a 3D-design bench whose output is the class 3D printer. Science labs are expensive because schools were told optics and motors must be bought — models made them nearly free.
The shadow class
Before a cikgu tries a new lesson on forty real children, it runs against forty simulated ones: student twins trained on the class's own anonymised mastery history, stress-testing the plan and predicting which three topics will buckle on Thursday. Rehearse on the twin — never on the child.
The weather map of what Malaysia knows
AnTIE fuses probe data from millions of cars into one honest state of the roads. The same fusion engine pointed at learning: millions of anonymised, aggregated probes become a live national topic map — which chapters the country is buckling under, grey wherever the data is thin. A ministry sees the storm two years before the cohort sits SPM.
Verifiable mastery
A transcript of what you proved, not what you attended. Every green tile is a signed attestation — the probe, the date, the standard it met — portable to polytechnics, apprenticeships, employers and national talent pipelines like the one PINTAR runs. A CV that cannot lie, because nobody typed it.
Companion minds with faces
The QynQue family already has them: Qynix the eagle strategist who plans the route, QuePo the purple hippo who sits with you at nine-thirty on a Sunday night. In AnTeraju they grow persistent, on-device memory of exactly what you struggle with — because a companion who notices you've avoided fractions for three weeks isn't decoration. It's trust.
Three horizons, in order
The sequence matters more than the ideas: the spine first, because everything flashy either projects from it or is theatre.
Ledger, gates, six surfaces
The part that cannot be faked later: one Learner ID, the append-only mastery ledger, the mastery gate, and the family's first surfaces. Everything below projects from this — nothing here ships before it does.
Machines that can read a page
Step-level error attribution, Socratic voice sessions, the five-year memory and Kampung STEM contexts — plus distilled tutor models small enough to run on the school's own QynQue box, with or without the line.
The weather map of what Malaysia knows
Shadow-class rehearsal, fused national topic state above strict aggregation thresholds, and verifiable mastery credentials that outlive the platform — the pipeline from a village classroom to a national scholarship, on provable ground.
The guardrails are the feature
A national map of what children know is the most sensitive dataset a country can hold. It is only buildable if the privacy was designed first — so it was.
Aggregate or nothing
The weather map is built from probe outcomes that clear a minimum-cohort threshold; below it the map stays grey by construction. There is no per-child row in a ministry view — the ledger's own scoping rules make one unrepresentable.
On-device by default
Reading a child's handwriting, listening to a revision explanation, holding five years of mistakes — this data lives on the family's or the school's machine. Sync is the exception the user chooses, not the default we collect.
Child-safe model contracts
Tutor minds are evaluated on abstention as a first-class metric: a mind that guesses at a Form 4 question fails our eval even when it is right 94% of the time. Jailbreak-resistance and no-ad-profiling ship in the same contract.
Concepts stay labelled
Every card above carries its true state — workshop or concept — in the same tradition as our status pages. We would rather under-promise a national graph than ship a dashboard that quietly invents the future.
Ready to build something intelligent?
Let's engineer your next product — from idea to launch.