Teaches the same concept five different ways depending on who is asking — without storing a single page of curriculum.

AI/ML

Engineering

Product

2025

Pinecone

RAG

Prompt Architecture

TypeScript

COPPA-aware

The interesting engineering here is restraint: personalize the delivery aggressively, but keep the vector database out of the curriculum business entirely.

The problem

Most AI tutoring demos pick one of two losing options. They flatten every learner into a single default voice, or they hoard curriculum content — taking on copyright exposure and a stale-content maintenance burden. The real challenge is adapting how a child is taught to their learning style while keeping what they are taught out of your storage, and making sure the personalization layer can never talk the tutor out of good pedagogy.

Solution & demo

A five-question assessment (information processing, pacing, help-seeking, motivation, confidence) bootstraps a learning-style profile for a hard cold-start problem. From there the tutor adapts tone, pacing, examples, and help-giving per learner — visual-spatial learners get diagram-first explanations, challenge-seekers get harder twists — while enforcing Socratic pedagogy for everyone.

Recorded walkthrough coming soon.

In the meantime, the architecture and tradeoffs below tell the full story — or start a conversation and I'll walk you through it.

Recorded walkthrough: pick a synthetic learner profile, ask the same math question, and watch the tutor's response change with the retrieved strategies shown beside it. Synthetic demo profiles only — no real child data anywhere near the portfolio.

Live “Try it” demo lands in a later release. For now, the recorded walkthrough above is the demo.

How it works

RAG used surgically: retrieve the top-5 teaching strategies for this child, not a curriculum dump.

For each turn, Pinecone retrieves the top-5 teaching strategies matching the child's traits, grade, and subject. Those strategies are injected into a four-block hierarchical prompt — Instructions → Context → Task → Output — with explicit priority ordering, so the learning-style layer can shape delivery but sits below the pedagogy layer. Lesson content itself is retrieved at runtime from open educational sources, never stored.

A profile-refinement loop keeps the model of the learner honest: observed session behavior — question frequency, response to visual hints, frustration signals — updates traits alongside assessment data, with per-trait confidence tracking and a graceful fallback to general strategies when a learner is new. Because the audience is children, data handling is COPPA-aware and safety-first by default.

Tradeoffs & lessons

What I'd do differently

Runtime retrieval vs stored curriculum. Fetching lesson content at runtime costs latency and adds a reliability dependency on external sources. Storing it would be faster and more predictable — and would import copyright liability plus a re-ingestion treadmill every time content changed. The scoping decision traded a little runtime risk for a lot less legal and maintenance risk.
Guardrails above personalization, always. Encoding the CLARIFY → ELICIT → NUDGE → SCAFFOLD → SOLVE hierarchy as a fixed prompt priority — rather than a suggestion — is what stops an adaptive tutor from degenerating into an answer key.
What I would do differently. Warm the cold-start faster by pre-clustering common learner archetypes, so the first few turns feel personalized before the refinement loop has enough signal.

Links

The demo above is the recorded walkthrough; a public repo and a longer write-up of the prompt architecture are being prepared and will appear here.