All work

Case Study

TalimAI

AI-planned computing curriculum for grades 5–12, with a teacher-in-the-loop gate

Self-directed build — product and system designWeb (SaaS)
EdTechAI CurriculumMulti-tenant

01Problem

Computing teachers for grades 5–12 spend hours planning differentiated daily lessons, and off-the-shelf content ignores what a specific class actually did yesterday. TalimAI generates a structured, grade-calibrated computing day for every class each night — built on what that class covered the day before.

02Constraints

  • Nothing AI-generated can reach a student unreviewed — a real classroom needs a human gate.
  • A “day” is not one quiz: it has to span warm-up, core skill blocks, an applied project, an optional challenge, and a wrap-up, calibrated to grade level.
  • Institutions, not individuals — the system has to model multiple campuses, teachers, and classes with real permissions.

03Approach

  • Generate a full 4–6 hour day nightly per class — warm-up → core skill blocks → applied project → optional challenge → wrap-up — calibrated to grade level and to the previous day’s coverage.
  • Support 9 distinct task formats (multiple choice / short answer, real code exercises, debug challenges, spot-the-bug / ordering, AI-concept design projects), not a single quiz type.
  • Close the loop: objective tasks auto-grade instantly; open-ended work gets AI-assisted scoring a teacher can review and override; results feed the next day’s plan before it is generated.
Product Thinking

The core product decision is a human gate: the AI plans, the teacher decides.

  • Every generated day is a draft — a teacher reviews and publishes it before any student sees it. Nothing reaches a student unreviewed.
  • Gamification layers on top of real academic content, never instead of it: XP, streaks with streak-freezes, deliberately class-scoped (not global) leaderboards, and a persistent per-student portfolio that carries artifacts forward day to day.
  • Teachers can upload and author their own content alongside the AI curriculum, so the tool augments them rather than replacing them.
  • Institution admins see real, itemized cost accounting for AI-generation spend — explicitly not a black-box number.

04Architecture

Architecture

Multi-tenant from the schema up — multi-campus, multi-teacher, multi-class, with granular per-teacher permissions — wrapped around a nightly AI generation pipeline and a grading loop that feeds the next day’s plan. (This was a self-directed build without a public repo, so the write-up stays at the level the product itself makes visible.)

  • Multi-tenant data model — campuses, teachers, classes, per-teacher permissions
  • Nightly AI generation pipeline — one calibrated day per class
  • Grading loop — auto-grade plus AI-assisted scoring with teacher override
  • Cost accounting for AI-generation spend, surfaced to institution admins

05Decisions & Tradeoffs

Made every generated day a draft behind a teacher publish-gate.

Adds a manual review step (and no fully-unattended flow on lower tiers), but nothing unreviewed can reach a student.

Scoped leaderboards to the class, not globally.

Less viral competition, but keeps motivation healthy and comparisons fair within a real cohort.

Surfaced itemized AI-generation cost to institution admins.

Exposes spend that some products hide, but makes the tool’s economics honest and defensible to a budget owner.

Gated access behind an application/review instead of self-serve signup, with tiers differentiated by manual vs. unattended nightly generation and roster ceiling.

Slower top-of-funnel, but fits an institution-sold product where onboarding is high-touch.

06Outcome

Shipped

  • A nightly-generated, grade-calibrated computing curriculum for grades 5–12 across 9 task formats.
  • A teacher review-and-publish gate, plus teacher-authored content alongside the AI curriculum.
  • An auto plus AI-assisted grading loop that informs the next day’s plan, gamification layered on real content, and itemized AI-cost accounting for admins.
  • Tiered plans (Starter / Growth / Enterprise) differentiated by manual vs. unattended nightly generation and roster ceiling.