Xamio is a full-stack adaptive learning platform built specifically for the AMIRNET — the standardized English proficiency exam used for university admissions across Israel. Every year, thousands of students prepare for it using generic English apps, static PDFs, or expensive tutors. None of them adapt to the individual.
Xamio does. It’s built on three proven mechanisms from learning science and psychometrics — Item Response Theory, SM-2 spaced repetition, and on-demand AI question generation — connected into a single system where every drill, every mock exam, and every vocabulary review feeds back into a unified model of the user’s ability.
The result: a platform where difficulty adjusts in real time, vocabulary schedules are personalised per word, and question supply never runs out. I scored 147/150 on the real exam. This is the system I built to make that result reproducible.
The AMIRNET is taken by every university-bound student in Israel. The prep ecosystem around it is underserved — there’s no Duolingo, no Magoosh, no Khan Academy for this specific exam. Students cobble together tools that weren’t designed for the format, miss the adaptive structure of the real test, and end up memorising answers instead of building ability.
Xamio fills that gap with a system grounded in the same psychometric frameworks used by the GRE and GMAT. It’s not a question bank with a UI — it’s a learning engine that models the student and optimises every minute of study time around that model.
The core insight is that three independent mechanisms from learning science — each proven in isolation — become significantly more powerful when connected into a feedback loop.
Most “adaptive” platforms use a simple if-else: correct answer → harder, wrong answer → easier. That’s not adaptivity. Xamio implements the 2-parameter logistic (2PL) model from Item Response Theory — the same framework behind the GRE, GMAT, and the real AMIRNET itself.
After every scored section, the system recalculates the user’s ability estimate (theta, θ) using maximum likelihood estimation. Theta ranges from -3 to +3 and maps to the AMIRNET’s 50–150 score scale via a nonlinear sigmoid function. Each question carries two IRT parameters:
This means the system doesn’t just know whether you got it right — it knows how informative that answer was. A high-discrimination question at your theta level tells the system more about your true ability than ten easy ones. The next section is calibrated to your updated theta, making difficulty selection continuous rather than discrete.
Vocabulary is the backbone of any language exam. The AMIRNET tests high-frequency academic English across 10 thematic domains — Science & Nature, Law & Ethics, Medicine & Health, Psychology, and more.
Cramming doesn’t work. The SM-2 algorithm (the same one behind SuperMemo and Anki) schedules reviews at increasing intervals, tuned to each word’s individual difficulty for the user. Every word in the 3,600+ word bank gets:
Rate a word “Easy” and you won’t see it for weeks. Rate it “Again” and it’s back tomorrow. The system learns your memory as precisely as it learns your ability.
Static question banks have a fundamental problem: they run out. Students who study long enough see the same questions, memorise answers, and get a false sense of readiness.
Xamio generates fresh questions on demand using NVIDIA Nemotron (purpose-tuned for academic English) with Groq-powered fallbacks for high throughput. Every question is:
All six AMIRNET formats are covered: Sentence Completion (1 or 2 blanks), Restatement, Reading Comprehension (passage + 5 questions), Grammar in Context, Word Formation, and Essay writing with AI evaluation.
The critical engineering challenge is latency masking. During a mock exam, the next section generates in the background while the user works on the current one. Sections 1–2 are pre-generated before the exam starts (theta = 0). Each subsequent section generates during the prior scored section, using the latest theta. The user never waits for AI.
Mock exams on Xamio aren’t question sets — they’re full adaptive simulations that mirror the real AMIRNET structure:
The exam evolves as you progress. Section 1 performance shapes section 2. Your theta carries across sessions, so the system knows your level from the moment you start.
After every session, the system analyses accuracy across every question type and domain, identifying consistent weakness patterns — not just one bad day. One tap starts a Weak Area Training session pulling from previously incorrect questions. Missed vocabulary feeds automatically into the SM-2 review queue.
The analytics dashboard surfaces: streak data with a 90-day activity heatmap, accuracy by type and difficulty, trending performance charts, due-word counts, and exam history with score trends. It’s the visible layer of everything underneath — the IRT model, SM-2 system, and weak area detection all surface in one place.
The hard part of Xamio isn’t any single feature — it’s the feedback loop between systems. IRT updates feed question generation, which feeds weak area detection, which feeds vocabulary review, which feeds back into performance metrics that update the IRT model. Making these work as one system instead of six disconnected features is what gives the platform its compounding value.
Every drill tightens the model. Every mock exam reveals what to study next. Every missed word enters a spaced repetition queue calibrated to your memory. The system doesn’t just test you — it learns with you.