Math & physics PDFs, end-to-end · verified, not vibed

Lessons that compile.
Math that checks out.

A single Rust binary that turns a topic prompt into a typeset, dimensionally-consistent .pdf lesson — Typst for rendering, all-Rust verification for correctness, OpenRouter for the words.

Open the dashboard Star on GitHub API

~ms: verification per exercise
0: external runtimes
1: static binary
7-vec: SI dimensional algebra

Why bother

The model writes. The runtime verifies.

Most "AI lesson generators" are wrappers that hope the model is smart enough. This one treats every lesson as a small DAG of jobs with mechanical gates between them — no reviewer model, no hand-curated facts file, no Python sidecar.

Typst, in-process

Real PDF compilation via the typst crate — vendored templates and embedded fonts, no shelling out, no LaTeX, no network at compile time.

All-Rust verification

Pratt parser + f64 evaluator. Every exercise gets numerically substituted at 16 deterministic samples — sign errors and lost factors die instantly.

Dimensional analysis

A 7-vector SI algebra walks every formula tagged dimensional: F = m/a fails before it ever reaches a reader.

Multi-stage pipeline

plan → fan-out write_section → compile-as-a-gate → verify → assemble. Each stage is its own job kind: retries, backoff, panic isolation come for free.

Queue groups

Network-bound LLM jobs and CPU-bound compile jobs live in separate worker pools, with a per-lesson fairness cap so one big lesson can’t starve the others.

Deterministic

Model, prompt version, seed, and raw stage I/O persist on every lesson. POST /lessons/:id/regenerate is bit-stable.

Pipeline

Five stages.
Two pools.

Each stage is a job kind. The job table is the orchestrator — durable, retryable, fair. LLM and CPU work live in separate pools so a slow OpenRouter call can never block a Typst compile.

llm #1

plan_lesson

outline + section scaffolding
llm #2

write_section

fan-out, one per section
cpu #3

compile_section

Typst hard gate
cpu #4

verify_section

numeric + dimensional
cpu #5

assemble_lesson

final PDF + .typ

Compile-as-a-gate

Each section's Typst is compiled before the section is allowed to merge. Compile errors feed the diagnostic back into a targeted retry — broken math never reaches the assembled PDF.

Wait for children

parent_id + children_remaining on the jobs table. Each completion decrements the parent in the same tx; when it hits zero the parent flips back to pending. No polling.

Per-lesson fairness

The claim query refuses to run more than N jobs that share a root_parent_id. Multiple lessons interleave; one big course can't starve the rest.

Verification

Schwartz–Zippel,
in microseconds.

Every exercise carries a verifiable answer. A small in-process Pratt evaluator substitutes both sides of the claimed identity at 16 deterministic random points — agreement at all 16 means equality with probability approaching one.

Numeric equality, polynomial identity, root-set verification.
Diff(expr, var, h) pseudo-function for derivatives via central differences.
7-vector SI dimensional analysis on every formula tagged dimensional: true.
All in-process. No sympy subprocess, no JIT, no network call per check.

verify_equal — d/dx sin(x) ≟ cos(x) PASS

Diff(sin x, x, 1e-6)

cos(x)

-3.14

+0.0000

≈ 0

-1.42

−0.9876

≈ 0

+0.27

+0.2666

≈ 0

+1.93

+0.9359

≈ 0

+2.71

+0.4187

≈ 0

… 11 more deterministic samples · seeded PCG64 · tolerance 1e-6

dimension check

F = m · a → kg·m·s⁻² vs N PASS

F = m / a → kg·s²·m⁻¹ vs N FAIL

outline validator

Every section declares defines and requires. A pure-Rust check rejects any outline where a section requires a symbol no earlier section defines — before a single writer call is made.

API

One POST.
One PDF.

Submit a topic, watch the pipeline drive itself through Server-Sent Events, download the rendered PDF when it's done. Regeneration is deterministic on the stored seed.

curl -X POST http://localhost:8080/lessons \
  -H 'content-type: application/json' \
  -d '{
    "topic": "derivatives via the limit definition",
    "level": "intro",
    "domain": "math",
    "lang": "en",
    "target_pages": 6
  }'

← 201 Created
{
  "id": "01J5K2A8...PTQ",
  "root_job_id": "01J5K2A8...XQZ"
}

also: GET /lessons GET /lessons/:id GET /lessons/:id/typ POST /lessons/:id/regenerate

Stack

No Python.
No LaTeX. No drama.

Everything ships as one Rust binary plus an SQLite file. The only network calls leave the box for OpenRouter and only when a writer needs a model.

Rust single static binary

Axum 0.8 HTTP + SSE

Typst in-process PDF render

SQLite durable job queue

OpenRouter DeepSeek V4 default

tokio queue-grouped workers

schemars JSON Schema repair loop

tracing structured logs + metrics

Boot the binary.
Watch a lesson typeset itself.

Open source, Apache + MIT, single binary. cargo run, point it at OpenRouter, then drive the pipeline from the dashboard or POST a topic.

Open the dashboard Read the source Quickstart

Lessons that compile. Math that checks out.