Personalized Learning at Scale

01. Problem

02. The AI digital textbook (math)

For a new math digital textbook, the goal was adaptive practice — the right problems for each student. The obvious tool, our team's deep knowledge-tracing model, needs a lot of per-student solving history that a brand-new product simply doesn't have yet. So I designed a rule-based system that works from day one and quietly accumulates the data to make ML possible later.

1. A knowledge state with no cold start

   Every sub-unit has a few common problems (3–5) every student solves, so there's never zero history. From correctness and item difficulty, the system estimates per-concept mastery.

2. Adaptive problem sets

   Based on that state, it assigns a tailored number and difficulty of problems, advancing through sub-unit review → unit → final assessment. Do well and it serves more, harder items; struggle and it serves more concept-reinforcing, easier ones — and a missed foundational item triggers a problem from the prerequisite concept in the knowledge map.

3. Signals beyond right and wrong

   Mastery rises and falls differently depending on item difficulty, and solving time counts too — rushing or guessing too fast is penalized rather than rewarded.

4. Built for teachers, in Java

   The deployment solution was Java-only: SQL pulls difficulty-appropriate items, Java assembles each problem set and caps repeats so the same problem doesn't keep reappearing. Teachers can also enter parameters (target difficulty, class size) to auto-generate a workbook from the problem bank.

Why rules, not a model? In education, teachers can't blindly trust a black-box estimate for high-stakes assessments — it has to be understandable and respect curriculum order. Rules earned that trust early, and I tuned them with field teachers, by grade level. Just as important, I designed the database to double as an ML foundation — logging problems served, correctness, difficulty, and achievement — so the product ships as rules today while accumulating the labeled history to move to ML (a Deep Knowledge Tracing–based diagnosis model I also researched) tomorrow. Delivered via RESTful APIs on the internal platform.

03. The learning platform at scale

On a corporate learning platform serving millions, the question flips from "what should this student practice next" to "which of thousands of courses would this learner actually want" — running live, across many tenants.

1. A two-layer recommender

   An LSTM over each learner's course and content history plus profile to predict what they'll engage with, and a BERT model matching their interests and role to course titles. Added cold-start handling (skipping learners and tenants with too little history) and filtering to drop mandatory and duplicate courses.

2. Production & MLOps

   Shipped it through Airflow — re-inferring daily on fresh activity and retraining weekly on the full history — with pre/post-processing to keep recommendations valid across a multi-tenant platform, served via RESTful APIs.

04. Results

• CVR and completion gains were measured over a 3-month post-launch monitoring period on a platform serving millions of learners.
• For the knowledge-tracing PoC, I also surfaced the model's explainability limits and proposed directions to make it interpretable enough for real deployment.

05. Beyond the platform

I shaped the AI approach for a national workforce-reskilling initiative — applying the same ideas (knowledge tracing, recommendation, and personalized coaching) to help people retrain for software roles as digital skills shift. Same conviction underneath: learning should meet people where they are.