Pass Mentors

One AI learning platform, many brands: generated courses, conversation coaching, examiner-grade feedback, and a readiness engine that says exactly when you're ready to book your exam.

Comment

PassMentors — Your AI coach from first lesson to exam day

One AI learning platform, many brands: generated courses, conversation coaching, examiner-grade feedback, and a readiness engine that says exactly when you're ready to book your exam.

Inspiration

Every exam has the same painful moment: "Am I actually ready to book?" Book too early and you burn hundreds of euros and your confidence. Book too late and you stall your career or your residence permit.

We saw it everywhere — newcomers to the Netherlands preparing for their inburgering and NT2 exams, students chasing IELTS bands, engineers sitting AWS certifications. Every tool on the market gives you content. Almost none of them will look at your data and tell you the one thing that matters: "Not yet — and here is exactly why, and here is the day you will be."

Tutors who can do that cost €40+ an hour. We wanted to build that tutor with AI — one that teaches you, talks with you, grades you like a real examiner, and is honest about when you're ready.

What it does

PassMentors is a full learning platform that runs multiple brands on one engine: PassMentors for professional certifications — AWS Solutions Architect (SAA-C03), Google Cloud Professional Cloud Architect, CompTIA Security+, CISSP, CISA, PMP, CFA Level 1, and the Bar (MBE) — Taalingo for Dutch and integration exams, plus English/IELTS and driving-theory storefronts. Everything a learner touches is AI-powered:

  • A certification coach with blueprint fidelity — across the whole IT and professional landscape. Cloud architecture (AWS Solutions Architect, Google Cloud Professional Cloud Architect), security and audit (CompTIA Security+, CISSP, CISA), project management (PMP), finance (CFA Level 1), law (Bar MBE): every track runs the same coached journey. A free 20-question diagnostic maps your strengths across the exam's domains in ~15 minutes → adaptive drills attack your weakest concepts → timed mock exams reproduce the real format → a readiness gate holds the booking recommendation until you clear the track's threshold. Each certification carries its own blueprint — its official domains, domain weights, question count, and time limit (AWS SAA, for example, drills four domains weighted 30/26/24/20 with 50-question, 75-minute mocks) — so practice always mirrors the exam you'll actually sit, whether that's IAM policies or the PMBOK. And because questions are generated per session against that blueprint, there's no static bank to memorize and no licensed bank to buy: a brand-new certification launches in days.

  • An automated content factory. Courses from absolute beginner (A0) to mastery are authored with LLMs, illustrated by image models that keep the same characters consistent across hundreds of lessons, and voiced line-by-line in studio-quality neural TTS — in the course's own language.

  • A study assistant that has read your course. Ask anything; answers are grounded in your actual lessons via RAG, with citations that link straight back to the source lesson. It indexes itself automatically on publish.

  • Live AI conversation. Every dialogue becomes a roleplay: the AI plays the other character at your CEFR level, then coaches you on what to improve. A realtime voice mode lets you speak with the AI partner in a phone-call-style session.

  • Examiner-grade feedback. Writing and speaking are graded asynchronously against exam-aware rubrics (IELTS bands, TOEFL scales, CEFR) — per-criterion scores with concrete feedback, in seconds instead of days.

  • Exams generated fresh, then verified. Every drill and mock exam is generated per session, targeted at your weakest concepts — and every question runs a verification gauntlet (the model re-solves it blind, a second model votes on the answer key, doc-grounded checks with citations) before a learner ever sees it.

  • The readiness engine — our wedge. A deterministic model (deliberately not an LLM) that tracks every concept on the official exam blueprint. It refuses to call you ready until you've proven breadth (90% concept coverage), dampens lucky streaks, detects when a domain is slipping, catches the answers you second-guessed from right to wrong, surfaces the domains you rush, and projects the date you'll hit the booking threshold. With certification exams costing $150–$700 per attempt, "not yet — here's why" is worth real money. It's the only coach that tells you not to book — and backs it with a pass guarantee.

  • Habit mechanics. Daily micro-learning with spaced repetition, tap-to-translate that feeds a personal word notebook, streaks, and an AI journal coach that corrects a few lines of daily writing with an explained correction table.

How we built it

One hexagonal backend, many brands. A FastAPI modular monolith (Python 3.12, SQLAlchemy async, Postgres) organized in bounded contexts — identity, catalog, content, learning, courseware, assistant, billing — with strict domain/application/infrastructure layering. Multi-tenant orgs and per-domain storefronts let PassMentors, Taalingo, and the English-exams brand share one engine with their own branding, catalogs, pricing, and locales (full UI in English, Dutch, and Arabic, including RTL).

The AI stack, each tool where it's strongest:

  • Question generation & authoring — LLMs generate questions per session against a concept × sub-skill taxonomy; a different model cross-verifies answer keys with majority voting.
  • Grading — exam-aware rubric prompts with strict JSON outputs, mapped onto an exam-agnostic 0–100 mastery scale so the readiness engine stays provider-neutral.
  • RAG assistant — course content chunked and embedded into Qdrant; retrieval is filtered per course and org, streamed over SSE with citations.
  • Realtime voice — Gemini Live native audio behind our own relay: the API mints short-lived signed tickets so the browser never holds cloud credentials, with per-minute credit metering.
  • Content media — Vertex image generation with a chained character-reference technique for visual consistency, and Chirp 3 HD voice-over with per-speaker voice casting; ~350 MB of generated audio across our courses.
  • Taxonomies from official sources — exam-guide PDFs go through Document AI, and every extracted concept must carry a verbatim source quote or it's dropped as fabricated.

The trust layer is classical code, not AI. Mastery EMAs with Bayesian-style shrinkage, coverage gates, regression and second-guess detection — fully deterministic and auditable, because a booking recommendation you can't explain is a booking recommendation you can't guarantee.

Everything runs in production on GCP with Celery workers for async AI jobs, Mollie for European payments (iDEAL), credit metering on every AI action, and an Expo mobile app sharing the same API.

Challenges we ran into

  • "AI-generated questions" is a liability, not a feature. One wrong answer key in a paid exam product destroys trust forever. We ended up treating generation as plumbing and building the real product around verification: blind re-solving, cross-model voting with confidence thresholds, doc-grounded checks, and post-publish answer-pattern anomaly detection.
  • Grading consistency across exams. IELTS wants half-bands, TOEFL wants 0–5 scales, CEFR wants descriptors. We solved it with exam-aware rubric adapters that all normalize into one internal mastery scale — so the readiness engine never needs to know which exam it's watching.
  • Realtime voice in the browser. Browsers can't attach cloud auth headers to live audio sockets. We built a ticket-relay pattern: the API signs a short-lived JWT, a thin proxy holds the only real credential, and billing is metered from the provider's own usage stream.
  • Character consistency in generated art. Naive image generation gives you a different "Sara" every lesson. Chaining a reference image through every generation call keeps one recognizable cast across a whole course.
  • Honest readiness needs honest data. Thin evidence must not look like mastery — we had to dampen small-sample confidence hard, and gate readiness on breadth of coverage, or three lucky answers made learners look exam-ready.

Accomplishments that we're proud of

  • A readiness signal we'd stake money on — literally: the model is conservative enough to support a pass guarantee, with certification records and a claims flow built in.
  • A content factory that shipped real catalogs: seven Dutch courses (A0–C2, NT2-style), six English levels plus IELTS masterclasses, three Business English courses — thousands of lessons, with generated illustrations and full voice-over, built in days instead of publisher-years.
  • Eight professional certification tracks on one engine, spanning five industries — cloud (AWS SAA, Google Cloud PCA), security & audit (Security+, CISSP, CISA), project management (PMP), finance (CFA Level 1), and law (Bar MBE) — each with a blueprint-faithful taxonomy extracted from its official exam guide, its own domain weights and mock-exam format, personalized drills, and the same honest readiness gate. Adding the next certification is a taxonomy import, not a content project.
  • The full learner loop is live in production — real domains, real payments, three UI languages, web + mobile — not a demo shell. AI usage is credit-metered per action, so the unit economics work.
  • Every AI claim in our demo is real: the citations, the roleplay, the band scores, the verification gates, and the readiness dial were all recorded from the working product.

What we learned

  • Verification is the product; generation is plumbing. Users don't buy "AI-generated questions" — they buy the confidence that the questions are right.
  • Don't use an LLM where you need to be accountable. The readiness decision is deterministic and explainable on purpose; LLMs feed it evidence, they don't make the call.
  • Grounding beats eloquence. Assistant answers with lesson citations and taxonomy entries with verbatim source quotes earn trust that fluent-but-unsourced answers never do.
  • Ship the whole loop. Content → practice → feedback → readiness → booking. Each AI feature alone is a gimmick; chained into one journey, they become a coach.

What's next for Pass Mentors

  • Roll out the Pass Guarantee publicly on our flagship tracks, backed by the readiness model's shadow-tested thresholds.
  • New exams and subjects, aiming for 100% coverage. The taxonomy extractor turns any official exam guide into a coached track in days — next up: more cloud and security certifications (Azure, AWS Developer/DevOps, CEH), Kubernetes, data and AI certs, more finance levels, and more languages and school subjects — until every high-stakes exam a learner faces has a track on the platform.
  • Voice everywhere: bring the realtime AI conversation partner from beta to every dialogue, including full spoken mock speaking exams.
  • B2B cohorts: team dashboards for language schools and employers (municipal integration programs, relocation agencies, cloud consultancies) with cohort readiness analytics.
  • Mobile parity to 100% and offline micro-learning for learners on the go.

Built With

Share this project:

Updates