8,500+ real ENT exam questions. Bayesian IRT in real time. Voice tutor in Kazakh. Multiplayer Battle. Built for Kazakhstan.
Where does the Krebs cycle take place inside a cell?
Trained on KazQAD·Built solo·Open to teachers and edtech marketers
Each layer is independently shippable, but together they replace static ENT prep apps.
01 · Adaptive engine
2PL Item Response Theory. After every answer, theta and standard error update via 41-point Gauss-Hermite quadrature. The next question is selected for maximum Fisher information at the current ability — same math behind SAT and GRE adaptive testing.
02 · Voice tutor
Whisper transcribes. Claude Sonnet streams the answer, grounded in 6,640 retrieved chunks from KazQAD and Kazakh Wikipedia. OpenAI TTS speaks back. Round-trip in ~3 seconds.
03 · Multiplayer Battle
Socket.IO matchmaking. 2-4 students answer the same 5 questions in real time. Speed bonus rewards quick correct answers. Winner gets +50 XP. A daily-active mechanism that competitors don't have.
Not vibes. Not roadmap. Already in production code today.
Real ML, not chatbot wrappers.
adaptive_engine 2PL Bayesian EAP · 41-pt Gauss-Hermite quadrature tutor_grounding Hybrid retrieval (vector + BM25) over 6,640 chunks voice_pipeline Whisper STT → Claude Sonnet (RAG) → OpenAI TTS-1 battle_runtime Socket.IO · NestJS gateway · 30s/question · matchmaking score_prediction Monte Carlo (2,000 samples) · per-subject θ posterior content_pipeline KazQAD ingest · Claude Haiku distractor gen · auto-validators
Static ENT prep apps (JUZ40, iTest) sell question banks. Examie sells outcomes.
| Feature | Static apps | Examie |
|---|---|---|
| Questions | Static bank | Bayesian-adaptive |
| AI tutor | None | Voice in Kazakh + RAG |
| Social | None | Multiplayer Battle |
| Content | Generic | KazQAD + Claude-augmented |
| Score prediction | Pass/fail | Monte Carlo CI |
| Pricing | ₸2,500–4,000/mo | TBD · parent-billed |
Citations, not hallucinations. Each layer is built on published work.
Dataset · KazQAD
Released by ISSAI / Nazarbayev University under CC BY-SA 4.0. 8,500 questions originally from real ENT exams, transformed and re-published. We license, ingest, and augment with Claude.
Method · 2PL IRT
Lord & Novick (1968). Used by ETS for the SAT and GRE adaptive sections. Our Python implementation runs a full Bayesian update in <1ms per answer.
Pattern · RAG
Lewis et al. (2020). Our hybrid retrieval (cosine + BM25) over halfvec(3072) embeddings keeps the tutor grounded — citations, not hallucinations.
Why not just use ChatGPT?
ChatGPT doesn't know the Kazakh ENT format, doesn't pull from real ENT questions, doesn't track your ability over time, doesn't speak Kazakh natively, doesn't have a competitive multiplayer mode, and won't score-predict your exam result.
Is the beta free?
Yes. Beta is free until Astana Hub residency confirms. Production pricing will be parent-billed (Apple IAP / Google Billing / Kaspi Pay), aimed at ~₸3,000/mo — well below the ₸15-30k/mo cost of a private tutor.
How accurate is the score prediction?
After 5–10 practice sessions, ±15 points 95% CI. After 50+ sessions, ±5–8 points. Monte Carlo over per-subject theta posterior. We re-calibrate as students complete real ENT mock exams.
What about minors and privacy?
Parent linking is mandatory for billing. Student PII is minimized — no school name, no real-time geo. KZ-hosted Postgres is on the roadmap. GDPR-compliant by design.
That’s how many take ENT every year. We’re starting with 50.
Request beta access