Case study

FlyingMinds

An always-on Learning Coach for grades K–10. I rebuilt the AI tutoring engine as sole engineer and took it from prototype to paying families.

1,500+
Students on the platform
$15K+
Revenue as sole engineer
81→92%
LLM grading agreement

The problem

Families want a tutor that adapts — not a worksheet factory. Early FlyingMinds had the vision (Fred as a patient Learning Coach) but the engine couldn’t diagnose mastery, route practice, or grade reliably enough to trust at scale.

What I shipped

I rebuilt the tutoring stack around Bayesian Knowledge Tracing for mastery state, a Node/Express service layer, DeepSeek for generation and grading, Postgres/Supabase for persistence (including pgvector where retrieval helped), Stripe for subscriptions, and an admin dashboard for ops.

On the product surface that meant adaptive testing, gamified vocabulary, parent-facing progress, SAT/ELA paths, and the Fred conversation loop: diagnose → adapt → advance.

Technical decisions that mattered

Mastery as state, not vibes. BKT forced every lesson and game to update a real skill estimate instead of “looks like they got it.” That made progress reports honest and routing deterministic.

Eval harness before prompt poetry. LLM grading disagreed with humans ~19% of the time. I built a labeled-set harness, measured agreement, and iterated prompts/rubrics until key agreement hit 92%. Shipping without that would have quietly corrupted mastery.

Model choice for unit economics. DeepSeek kept interactive tutoring affordable at $29/month while still good enough for explanations and grading under the harness.

Sole-engineer scope control. Payments, admin, and the tutor loop shipped first. Nice-to-have curriculum chrome waited. Revenue and retention told me what to build next.

Architecture (interview version)

Client (Fred UI) → API (session, diagnosis, practice routing) → mastery store (BKT skill params) → LLM layer (explain / grade / generate) → eval harness offline → Stripe + admin for business ops. Failures in grading feed the harness; harness wins become prompt/rubric changes, not one-off hotfixes.

What I’d tell a hiring manager

FlyingMinds is the clearest proof I can own an AI product end-to-end: ambiguous learning science, LLM reliability, payments, and user-facing craft. I’m keeping it alive as a product — not as a substitute for joining a team where shipping is the job.

Visit FlyingMinds Interview kit