Grade Coach

The AI grading tool teachers trust. Snap photos of student work, grade the whole stack, review the results, export the results. You're still the teacher... just more smarter, more better, more faster.

Comment

Inspiration

I taught for over a decade in international schools - RMIT, Scotch AGS, ISB, HCMC University of Science. The grading pile was killing my weekends. I tried the existing AI grading tools and ran into the same wall every time: they were either chatbots pretending to grade, or enterprise platforms with a six-month district sales cycle. Nothing for the working teacher with 100 papers on a Sunday night.

The wedge insight: teachers do not want a chatbot. They want their own rubric applied consistently to the whole stack. So I built that.

What it does

Grade Coach is an AI grading tool for K-12 teachers:

  1. Snap photos of student work (or upload scans, PDFs, files from a phone).
  2. The AI reads the first paper and proposes a rubric. The teacher reviews + locks it.
  3. With the rubric locked, the AI grades the rest of the stack the same way.
  4. The teacher reviews every score, edits anything, exports to CSV / Google Sheets / PDF.

Seven tools share a per-student concept-mastery substrate that compounds over time:

  • Grade Papers - the core grading flow
  • Worksheet Studio - printable A4 worksheets from any reference
  • Parent Reports - per-student personalized parent comms
  • Student Views - read-only feedback pages per student
  • Practice Generator - per-student practice worksheets from substrate-detected weak concepts
  • Personalized Grading - any number of papers, each graded against its own answer key
  • Class Deep-Dive - class narrative + concept heatmap + mastery timeline

The substrate is the moat. Grade Coach is not just faster, it remembers.

How I built it

The product is AI-native end to end:

  • Grading runs on Gemini 3.1 Flash-Lite for speed (Netlify functions cap at 26 seconds, so Pro was too slow). Each paper is one Gemini call; rubric-lock is a separate call against the first paper.
  • Worksheet design, parent reports, student views, and class deep-dive narratives run on Gemini 3.1 Pro - the longer-form reasoning calls where quality matters more than wall-clock.
  • Worksheet rendering is a two-step pipeline: Gemini Pro for the editable design brief, then gpt-image-2 for the printable A4 image (a background Netlify function so it can exceed 26 s).
  • Per-student substrate writeback fires on every grade pass: the analyze call emits Bloom-level, CEFR, and language-load tags per question, persisted to Postgres so future tools can read mastery over time.
  • 20 Gemini API keys rotate to parallelize grading without hitting rate limits - fan-out gives roughly 6x wall-clock improvement on a full stack.

Stack: React 18 + TypeScript + Vite + Tailwind v4 + Zustand + React Router v7 on the front end; Netlify Functions + Supabase Postgres on the back end; Stripe for payments; PostHog for analytics; full i18n in English / Spanish / Vietnamese. Continuous deploy from main on every push.

Challenges I ran into

  • Netlify's 26-second function timeout. Forced Flash-Lite for grading. Parallelized via Gemini key rotation to claw back the wall-clock.
  • Substrate normalization. Different teachers tag the same concept differently. Pgvector + embeddings + LLM canonicalization is wired up but kept dormant until the concept count per teacher hits the threshold where it actually matters.
  • Stripe webhook race condition. Two simultaneous deliveries of a top-up purchase could zero out one. Fixed with an atomic SECURITY DEFINER RPC and a stripe_events claim-then-handle pattern.
  • HMAC grade-receipts. Built so anonymous users (no account) can grade from the free tier without exposing the AI endpoints to scraping. Receipt is signed server-side, roughly 10-minute TTL, consumed once.

What I learned

  • Sequence the moat. Grading was the wedge. The per-student concept substrate is the moat. Every Phase 3 tool is a surface onto the substrate, not a standalone feature. That sequencing decision was hard but right.
  • Build for the wedge persona. International-school ESL teachers are who I know. Designing for them - rubric-first, photo-first, classroom-language-tolerant - beats trying to please everyone.
  • AI is the operator, not a feature. The hardest engineering was not "calling an AI." It was making the AI the production system: grading on every paper, parent reports written by AI, deep-dive narratives written by AI, worksheets composed by AI. Humans are the final check, not the producer.

What is next

  • Referral credits (designed; +25 free credits per qualified referral).
  • Substrate normalization activation (pgvector wired, dormant; flips on at scale).
  • Pipeline grading: start grading on the first paper drop, do not wait for the full stack.

Built With

  • gemini-3.1-flash-lite
  • gemini-3.1-pro
  • gemini-api
  • gpt-image-2
  • netlify
  • netlify-functions
  • postgresql
  • posthog
  • react
  • react-i18next
  • react-router
  • stripe
  • supabase
  • tailwindcss
  • typescript
  • vite
  • zustand
Share this project:

Updates