TradeTEST.TRAINING — Shipped AI SaaS

Live: https://tradetest.training · Code sample: 02_ai-engineering-code/tradetest-lean/api-explain-back/route.ts

Problem. California tradespeople need to pass the CSLB Class B contractor exam (Law & Business + Trade, 72% to pass) and study on their phones between jobs.

What I built. A full-stack, revenue-generating SaaS: 131 lessons / 13 modules, 1,500+ practice questions, 3-tier pricing (Free / $49 / $149), bilingual EN+ES.

Stack. Next.js 14 (App Router) · TypeScript (strict) · Supabase Postgres + Auth · Stripe · Groq (Llama 3.3 70B) · Resend · PostHog · next-intl · Playwright · Vercel.

AI engineering.

Explain-back tutoring — a Socratic LLM endpoint that evaluates a learner's own explanation, runs 2–3 coaching turns, and wraps up; system prompts tuned to an 8th-grade reading level.
SM-2 spaced repetition drives adaptive question selection.
Tier-gated LLM spend — AI features are budgeted per pricing tier.

Why it matters for an applied-AI role. Proof I can ship an LLM feature inside a real product with auth, billing, i18n, and tests around it — not a demo. Production prompt design + cost control.

Talk track (screen-share)

This is live in production at tradetest.training, paying customers on it today. The piece I want to show you is the explain-back endpoint: instead of grading a multiple-choice answer, I ask the learner to explain the concept back, then send that to Groq to evaluate and run two or three Socratic follow-ups. The interesting engineering isn't the prompt, it's the guardrails around it: LLM features are gated by pricing tier so spend is bounded, and everything sits behind Stripe, Supabase auth, and a Playwright suite. I treat the model as one component in a product, not the product.

GG · Applied AI Engineering portfolio · synced to youdontneedmy.help design