AUTONOMOUS REEL STUDIO — MANIM + LLM + TTS

Generate a synced motion reel from one prompt.

Type “How Redis pub/sub works” or “Raft Consensus” — Shawtty researches, scripts, narrates, animates in 1080×1920, and delivers an MP4 with music and voice in sync.

20s · 9:16

Watch sample output live studio · shawtty.spacebuilder.in

1080×1920 9:16 HD
Sarvam AI Voice
Self-Healing Renderer
Docker One-Command Deploy

shubh · bulbul:v4 0:20

job_7f3a2c complete

How a voice AI agent helps a dental clinic answer missed calls, book appointments, and send reminders.

✓ Searching web
✓ Planning motion graphics
✓ Generating narration
✓ Rendering 1080p Manim reel
✓ Mixing music/SFX

Creativity scores

hook

9.1

variety

8.4

motion

8.8

sound

8.2

originality

8.6

Video Audio

01 / PIPELINE

Five passes from a sentence to an MP4.

Each stage writes an artifact you can inspect. Nothing is a black box — the trace log, beats table and probe frame are all in the Studio.

STEP 1
Plan 4-Act Storyboard

LLM plans hook, mechanism, insight, takeaway.
→ plan.json
STEP 2
Synthesize Voiceover

Sarvam AI generates per-act audio and measures its duration.
→ act_1.mp3 · 4.82s
STEP 3
Generate Time-Synced Code

LLM writes a self-contained ReelScene Manim class paced to the audio.
→ reel_scene.py
STEP 4
Render 1080p + Self-Heal

manim render -qh at 1080×1920. Auto-repairs on error.
→ 1080×1920 · 60fps
STEP 5
Mux & Deliver

FFmpeg mixes narration + SFX + video, then streams to your browser.
→ output/job_id.mp4

02 / NARRATIVE PLANNER

Every topic gets the same four acts, timed to the voice.

claude-opus-4.6 via LiteLLM breaks the prompt into a story arc. Sarvam measures each act to the millisecond, and every Manim run_time is aligned to those beats.

plan length · requested 20s · actual 20.0s voice: shubh_enhi_ads · bulbul:v4-flash

ACT 1 · 4.2s Hook & Problem “Your clinic misses 1 in 4 calls.”

ACT 2 · 7.4s Mechanism Deep Dive “The agent answers, checks calendar, books.”

ACT 3 · 4.6s Aha! Insight “Every missed call becomes a booked slot.”

ACT 4 · 3.8s Real-World Impact “Reminders cut no-shows overnight.”

0s2s4s6s8s10s12s14s16s18s20s

Voice-locked beats

Word timestamps from the narration become animation cues. Entrances fire on atWord, SFX land on scene starts.

start	end	word	cue
0.00	0.42	Your	Write(title)
0.42	0.91	clinic	FadeIn(icon:user)
0.91	1.38	misses	—
1.38	2.06	one in four	Transform(counter)
2.06	3.20	calls.	Flash(phone) + sfx:pop

03 / UNDER THE HOOD

A real render engine, not a template library.

SELF-HEALING

The repair agent reads the traceback, so you don't.

If a Manim script fails, the agent reads the Python traceback and stderr, patches the code and recompiles — up to 3 retries. Jobs rarely fail; when they do, you get a plain-language reason.

attempt 1/3

TypeError: Arrow.__init__() got an

unexpected keyword 'tip_len'

repair → reel_scene.py:42

- Arrow(a, b, tip_len=0.2)

+ Arrow(a, b, tip_length=0.2)

attempt 2/3 · rendered 1080×1920 ✓

PROCEDURAL MIX

Music ducks. Bottom 20% stays clear.

FFmpeg sidechain compression ducks music under voice; SFX hit on scene starts. No text in the bottom 20% — safe for Instagram UI.

SARVAM AI VOICEOVER

Narration measured to the millisecond.

bulbul:v4-flash with shubh_enhi_ads voice profile. Scene run_time auto-aligns to the audio beats with edge-tts fallback.

PURE MANIM COMMUNITY

Vector type, real geometry, 60fps.

Crisp typography, smooth geometric transforms, dynamic graphs and architecture diagrams — rendered natively at 1080×1920.

☁️

🖥️

🗄️

⚡

🔒

SVG ICON LIBRARY

Bundled technical icons.

Cloud, server, database, redis, cpu, lock, user, activity, networking, security.

🖥️

DUAL INTERFACE

Studio for output. Console for code.

Web Studio :3000 for prompts, job tracking and reel viewer. Manim Service :8000 for live code inspection.

🐳

CONTAINERIZED

One command to run all of it.

docker compose up --build ships Python 3.14/3.11, Manim, FFmpeg, Cairo, Pango and Node.js.

04 / EXAMPLES

Topics people actually explain.

Distributed systems, API auth flows, AI agents for business owners, clinic automation, dev-education Shorts. Pick one to load it into the composer.

DISTRIBUTED SYSTEMS

Redis pub/sub vs message queues with worker pools

CONSENSUS

Understanding Raft Consensus in distributed databases

API AUTH

OAuth 2.0 Authorization Code flow with PKCE

NETWORKING

How WebSockets work under the hood

AI FOR BUSINESS

How a voice AI agent helps a dental clinic book appointments

05 / THE STUDIO

Watch every job, down to the frame.

Prompt composer, run panel with progress ring, steps, plan length, unsupported list and the reel player.

localhost:3000

Prompt

How Redis pub/sub works vs message queues with worker pools

Duration

20s

Voice

shubh

Quality

1080p · 60fps

Launch Full Studio →

70%

Rendering 1080p Manim reel

Phase 4 · compiler active

1080×1920

06 / ARCHITECTURE

Two services, one compose file.

User / Browser

→

Web Studio

:3000

→

Express Gateway

server/index.ts

→

MANIM SERVICE · :8000

Orchestrator → LLM Planner → Sarvam TTS → Repair Agent → FFmpeg Mux

Frontend / Gateway	Node.js 22 · Express 5.1.0 · TypeScript · Better Auth
Engine	Python 3.14 · FastAPI · Manim Community · FFmpeg · Cairo · Pango
AI & Speech	LiteLLM Gateway (claude-opus-4.6) · Sarvam AI bulbul:v4-flash
Auth & Mail	Better Auth · Resend API (no-reply@luqe.in) · SQLite (node:sqlite)