1 of 10

Chalk

One prompt. A lesson you can follow.

Agentic teaching decisions. Deterministic visual execution.

Presented by Avinash Yadav

2 of 10

Most AI video optimizes for spectacle.

Learning needs structure.

A useful lesson must sequence ideas, keep labels readable, reveal relationships at the right moment, and remain clear after the wow factor fades.

3 of 10

ONE PROMPT IN — A COMPLETE LESSON OUT

A question becomes a lesson.

“How does the water cycle work?” → script, diagrams, narration, captions, and MP4.

▶ LIVE PRODUCT

Submit a topic. Watch progress, HLS preview, and the completed lesson.

4 of 10

END-TO-END ARCHITECTURE

One prompt travels through

a typed, observable pipeline.

1 PLAN · Gemini + TTS

2 COMPILE · TypeScript + Zod

3 DELIVER · Convex + FFmpeg

Next.js

React studio

Convex

queue + state

Node worker

claim + lease

Gemini + TTS

teach + narrate

Storyboard

sanitize + Zod

Compiler

layout · SVG · Sharp

FFmpeg

HLS · MP4 · VTT

5 of 10

KEY ENGINEERING DECISION

Agents explore.

The compiler enforces.

Planning agents never place the final video pixels. They emit bounded JSON; TypeScript owns geometry, timing, quality checks, and export.

Sanitize → Zod typed boundary

AGENTIC

• Research + outline

• Script editing

• Visual direction

• Scene design + icons

DETERMINISTIC

• JSON sanitizers

• Zod storyboard

• Layout + collision repair

• Timing + encoding

TYPED

CONTRACT

6 of 10

RENDER + MEDIA PIPELINE

Speech creates the timeline —

code produces every frame.

Per-beat Google TTS clips → measured PCM durations → interpolated reveal cues → SVG → frame-hash dedupe → Sharp → FFmpeg → HLS + MP4 + WebVTT

1 measured beat audio

2 SVG + Sharp frames

3 HLS + MP4 + WebVTT

The media pipeline is driven by speech. Google Cloud Text-to-Speech synthesizes one clip for each narration beat, and Chalk measures the PCM duration to place every beat at an exact offset. Phrase-level reveals inside a beat are interpolated from narration cues rather than claiming word-level TTS timestamps. Those same compiled beat boundaries produce captions and WebVTT. The renderer generates deterministic SVG, hashes frames to skip unchanged visual holds, and uses Sharp only for unique PNGs. FFmpeg emits sequential HLS segments after a short runway and creates the final H.264 and AAC MP4, whose duration is probed and validated.

[Sources]

- F:\wbev\worker\google.ts — per-beat synthesis and measured durations

- F:\wbev\worker\render.ts — SVG, frame dedupe, Sharp, HLS, and MP4

[/Sources]

7 of 10

RELIABILITY + SCALE

Every render owns a leased attempt.

1 → N

workers, one Convex queue

CLAIM

authenticated + atomic

LEASE

attempt ID + expiry

RENDER

renew + upload

COMMIT

current attempt only

Stale callbacks rejected · cron marks silent attempts failed · owner Resume requeues

Each render is tied to an authenticated attempt and lease. A worker atomically claims a queued job from Convex, receives the current attempt identifier and expiry, then renews that lease through heartbeats, progress, HLS, and uploads. Every callback must match the worker, attempt, and live lease; late workers cannot overwrite a newer attempt. A minute cron marks silent attempts failed in bounded batches, and the owner can Resume to requeue the job, potentially reusing that worker's local checkpoints. Additional outbound-only workers increase throughput, one video per worker, while replica count remains manually configured today.

[Sources]

- F:\wbev\convex\jobs.ts — claims, leases, callback validation, and stalled-attempt handling

- F:\wbev\DEPLOY.md — worker replicas and capacity model

[/Sources]

8 of 10

POLISHED FROM PROMPT TO MP4

A complete, accessible user flow.

Responsive studio

Desktop and 390 px mobile layouts.

Accessible playback

Keyboard focus, captions, transcript, live status.

Complete workflow

History, gallery, HLS preview, MP4 download.

Private by default

Owner-only regenerate/delete + custom dialog.

Designed for keyboard, screen-reader, mobile, and completed-state usability.

9 of 10

ENGINEERING EVIDENCE

Measured, modular, and release-ready.

284 / 284 passed

Deterministic unit, integration, regression, and edge cases.

3 / 3 Convex

Ownership, capacity, stale-attempt integration tests.

4 / 4 browser

Desktop + mobile Playwright scenarios with axe.

86.69% lines

86.42% branches · 83.47% functions.

Typecheck · lint · format · production build · Docker/CI · deployment docs

This is not only a feature demo. The repository separates the Next.js interface, Convex functions, worker, shared compiler, scripts, assets, and benchmarks. Zod and TypeScript protect the scene and API boundaries. The verified release passed 284 of 284 deterministic tests, 3 of 3 Convex integration tests, and 4 of 4 desktop/mobile browser scenarios with axe. Instrumented coverage measured 86.69 percent of lines, 86.42 percent of branches, and 83.47 percent of functions. Lint, formatting, both type checks, the production build, CI, and the Docker worker build also pass, with README and deployment documentation for repeatability.

[Sources]

- F:\wbev\README.md — Repository map, Quality checks, and release verification

- F:\wbev\package.json — test and release scripts

- F:\wbev\.github\workflows\ci.yml — CI and Docker jobs

- F:\wbev\DEPLOY.md — deployment, smoke-test, rollback, and judge-day checklist

[/Sources]

10 of 10

One question → a lesson anyone can follow.

Built for learning. Engineered for trust.

Live demo

https://chalkai-omega.vercel.app/