Case study

ClassFoundry: a working education platform, shipped and still shipping.

ClassFoundry is the live product behind Aesir's four-week-cycle claim. The four-week cycle is the unit of delivery, not the lifetime of the product. ClassFoundry's first working version went live in two weeks, and it has shipped on that cadence every month since.

Dec 2025first version live 310+commits in ~5 months 777automated backend tests Most weeksa new release ships

I. What it does

From a class to parent-ready feedback.

ClassFoundry helps independent online teachers turn a class into per-student feedback and parent-ready reports without doing it by hand. This is the loop a teacher actually runs.

  1. Record the class live, or upload a transcript.
  2. ClassFoundry drafts feedback for each student, informed by that student's history.
  3. The teacher reviews, edits, and accepts. Nothing is sent on the model's say-so.
  4. Export a parent-ready report.
ClassFoundry screen for starting a feedback run from a transcript or recording
Start a feedback run from a transcript or a live recording. Student details obscured for privacy.
ClassFoundry per-student feedback review screen
Per-student feedback the teacher reviews and accepts. Student details obscured for privacy.
ClassFoundry parent-ready report
The parent-ready report at the end of the loop. Student details obscured for privacy.

Beyond that loop, ClassFoundry records live classes through a meeting bot, grades assignments, keeps a reusable library of lesson materials, runs subscriptions and billing through Stripe, and ships a full Simplified-Chinese interface for teachers working across borders.

II. How fast it shipped

A working slice in two weeks, then a steady cadence.

ClassFoundry started as a clickable prototype on December 15, 2025. Within two weeks, by December 31, the working slice was live: accounts, billing, student and class management, AI feedback that remembers each student, a Simplified-Chinese interface, and a full UI pass.

The pace held. Across roughly five months the project has more than 310 commits, 40 or more every month, with live recording, assignment grading, the lesson library, and report export all added on that steady cadence. A public changelog records what shipped and when. 777 automated backend tests, plus a frontend suite, keep the cadence from breaking what already works.

  • Dec 15, 2025 Clickable prototype approved (Phase 0).
  • By Dec 31, 2025 Working slice live: accounts, billing, students and classes, AI feedback with memory, Simplified-Chinese interface, full UI pass.
  • January to May 2026 40 or more commits a month: live recording, assignment grading, lesson library, report export.

That is the part that matters for a prime. When a tool has to be live before a cohort starts or before a performance period closes, the cadence is the deliverable.

III. How it's built

One well-factored full-stack application.

ClassFoundry is a single full-stack application, which is part of how a small team keeps shipping weekly.

  • Frontend. React 18 and Vite, with Material UI.
  • Backend. Flask 3 on Python 3.12, SQLAlchemy with versioned Alembic migrations.
  • Data. PostgreSQL in production, SQLite for local development and tests.
  • Hosting. Railway.

Outside services each sit behind their own layer, so any one of them can change without touching the rest: Anthropic's Claude generates the feedback, Stripe handles billing, Recall.ai records live classes, and Resend sends mail. Incoming webhooks are signature-verified and deduplicated before they touch the database.

IV. How it's secured

Controls built into the product.

These are controls implemented in the codebase, not the result of a formal audit. They reflect how the product is built.

  • Authentication. JWTs in HTTP-only cookies, bcrypt-hashed passwords, and decorators that gate protected routes on login, verified email, and active subscription.
  • AI safety. Untrusted content such as student names and transcripts is wrapped in a nonce-based boundary before it reaches the model, to resist prompt injection.
  • Webhooks. Incoming Stripe and Recall.ai webhooks are signature-verified before processing. Recall.ai events are deduplicated against a durable event log; Stripe handlers are idempotent and validate customer ownership before applying a change.
  • Data isolation. Every query is scoped to the user who owns the data, and errors return a generic message while the detail stays server-side.

V. For a prime

The same model fits a subcontract.

The operating model behind ClassFoundry is the one Aesir brings to a subcontract: a working tool delivered inside a performance period or before a pilot cohort, built and revised on a weekly cadence, with the integration and security practices above already in hand. Aesir works under your methodology and your contract. You keep the field relationship and the facilitation margin.

VI. Contact

See whether this fits your next pilot or performance period.

A 30-minute call is the right first step. We'll talk through what's on the calendar and whether a four-week build is worth scoping.

Book a 30-minute call