An AI tutoring platform that witnesses real learning — and verifies it's real. Designed, directed, and validated end-to-end.
A production, multi-tenant adaptive-learning SaaS I designed, architected, and directed (AI as the implementer) over about three months, solo. MAP-driven, BKT-adaptive instruction across eight domains, bank-first so it can't hallucinate facts — and an academic-integrity layer that catches AI-pasted work and makes a student prove she understood it. Here's how it's built, and the calls I made.
Everything shown is a 100% synthetic student— “April Brown,” a fictional grade-11 competitive climber — on a throwaway database. No real child's name, scores, essays, or private data appears anywhere on this page. The product is real; the person is invented. It's a ~3-month solo build, dogfooded — this shows capability and judgment, not traction.
Homeschool parents spend thousands on curriculum but still can't tell if their child truly mastered the material — and now there's a sharper fear: a kid uses AI to produce perfect work and learns nothing. When it's time to apply to college, years of real learning are scattered and have to be rebuilt into a record from scratch. The hard part isn't generating lessons — it's proving the learning is real and earning a parent's trust in what an AI puts in front of them. IvyBloom is built around that boundary: measure where the student actually is, generate from verified banks so facts can't be invented, and let nothing count as mastery until the student proves the work is her own.
Witness real growth. Verify real learning.
The platform is organized around one idea: read the student's true arc over time, generate only what can be trusted, and let nothing count as mastery until it's verified the student's own work. The AI is the research assistant; the parent stays the author of record.
How the system works — and why it's built this way.
A production, multi-tenant AI homeschool-tutoring platform. I designed the whole thing: MAP-driven, BKT-adaptive instruction across eight domains, a bank-first anti-hallucination pipeline, an academic-integrity layer, parent-billing, and a full college-admissions suite.
Most college counseling sees a single polished senior year. IvyBloom reads the student's learning over time — three MAP test points showing +10 / +9 / +9 RIT growth— and turns that longitudinal data into the next concrete action. I made difficulty MAP-driven and Bayesian on purpose, so it comes from real assessment data, not a guess.
A parent's real fear: a kid uses AI to hand in perfect work and learns nothing. So I designed an integrity layer. When a submission reads above the student's level, it's flagged and won't count until she proves she understood it— a comprehension check, not a black-box accusation. A strong, honest student still scores 89, not 100, because the guardrail actually fired.
It generates a full week from the student's level and weak concepts — but bank-first: verified question banks for math and science, the LLM only where no bank exists, and even then validated. Anti-hallucination by construction, because one wrong math answer loses a parent forever.
It generates admissions material — an executive synthesis, an ability narrative, an evidence portfolio — but every generator is labeled “raw material, rewrite in your own voice.”The system is the research assistant; the parent is the author of record. And the synthesis names gaps, not just strengths — an evaluator that only praises is useless.
Underneath is a real three-year transcript I can issue as an immutable, serial-numbered record— and a public link an admissions office opens with no account, that I can revoke. Plus a College List that auto-buckets Reach / Match / Safety from real SAT and GPA, an essay coach, and a live FBLA business simulation where an AI investor probes the economics and pushes back.
It's not a prototype: tiered billing charged to the parent, role-based admin, God-Mode overrides, and per-call AI cost tracking. I audited the billing-bypass paths before building features — in a paid product, correctness is a security property.
The four calls that define the system.
Academic-integrity layer
Stylometry, readability, and behavioral signals catch AI-pasted and rushed work. A flag routes to a comprehension check, not an auto-accusation — passing clears the penalty without erasing the record. Product-safety thinking specific to an education product.
Bank-first anti-hallucination
Verified question banks are preferred over the model; the LLM only fills gaps where no bank exists, and its output is validated against a curriculum concept list. Anti-hallucination by construction — judgment about where not to trust a model.
MAP-driven + BKT adaptive engine
Difficulty comes from real NWEA MAP scores; weak-concept targeting comes from Bayesian mastery tracking. Real adaptive-learning modeling — not a quiz app with a difficulty slider.
Parent-billing + safety rails
Student actions charge the parent; God-Mode and impersonation skip billing; rate limits and refund-on-failure protect the paid AI routes. Financial-grade correctness in a multi-tenant system, audited before features shipped.
- 01Designed & architectedthe platform — the MAP-driven + BKT adaptive engine, the bank-first anti-hallucination pipeline, the academic-integrity layer, and the admissions suite.
- 02Directedthe AI build — wrote the specs, set where the model may and may not be trusted, reviewed the output, and decided what shipped. (Halted a finished feature when AP-level chat would have overloaded absolute beginners — domain judgment over code.)
- 03Validatedit end-to-end — audited the billing-bypass paths before features shipped, caught the logic bug that wrongly auto-stopped an athlete for switching to another study tool, and QA'd on the real running app, not “it compiles.”
This is how I work: design the system, draw the trust boundary, own the verification.
If you're evaluating someone to design or operate AI-augmented systems, this is a representative piece of how I think. There's more in the collection.