📖 A source-to-syllabus course

Wera: Anatomy of an Application OS

Wera is a real, deployed internship-application operating system — it discovers fresh postings overnight, tailors LaTeX resumes with AI that is structurally forbidden from inventing experience, compiles PDFs in a sandbox, and keeps its user applying with a momentum scoreboard. This course walks the codebase in the exact order it was built. By the end you'll be able to read a full-stack TypeScript monorepo top-down, explain every architectural decision it made — and argue with the ones you'd have made differently.

TL;DR

Seven modules, each in two depths: an Overview (plain-English intuition + the build journey) and a Deep Dive (complexity, tradeoffs, failure modes, exercises). Read Overviews for the story; read Deep Dives to level up as an engineer. The codebase analysis and architecture map are your reference pages.

🧭
How to read this course

Modules follow the real implementation order — spec first, riskiest link second, data model third. Concepts marked Simply / Technically can be flipped between intuition and rigor. Every technical term has a dashed underline: hover or tap it for a plain-English definition.

01

The Product & The Loop

What Wera is, why it exists, and the daily loop it optimizes. How a monorepo of one app and five packages maps to the product — and why the spec was adversarially reviewed before a single line of code.

02

Schema & Auth Foundations

Twenty-three Postgres tables as the product's nouns. Multi-tenancy from day one, an auth vendor kept behind a seam, and constraints that make invalid states unrepresentable.

03

The AI Boundary

A customs checkpoint between the app and any language model: schema validation, one corrective retry, a monthly budget brake, and a ledger of every call. The app never sees raw model output.

04

Evidence-Gated Tailoring & the LaTeX Sandbox

The differentiator: every AI-written resume bullet must cite real profile evidence or it never reaches the screen. Then the compile pipeline — running untrusted LaTeX like a stranger's code, with no secrets, no network, and a kill switch.

05

Discovery: Connectors, Quarantine & Dedup

Six job sources behind one typed interface. Items are guilty until validated — parse failures get quarantined, never dropped. Fingerprints collapse the same job seen on three sites into one row with three receipts.

06

Workflow Surfaces: State, Tasks & Time

A ten-status state machine that only moves forward, needs-flags that can't lie because they're computed instead of stored, idempotent "mark applied", and an hourly reminder courier that never delivers twice.

07

The Instrument: Design System as Code

"Precision instrument on warm paper." One accent color rationed to momentum only, a motion vocabulary with eight entries, and a coach line generated by a pure function — design decisions enforced like engineering decisions.

💡
Why this codebase is worth studying

Wera is small enough to hold in your head (~12,000 lines) but production-real: deployed, cron-driven, multi-vendor, and battle-scarred. Its git history contains actual bugs, actual review findings, and actual reversals — the things tutorials sand off.