Portfolio project · AI automation

Media Factory: a multi-agent AI studio for short-form video

An end-to-end system that finds source material, writes and fact-checks a script, voices it, edits and renders a vertical video, runs automatic quality checks and publishes on a safe schedule. A team of specialised AI agents operates and improves it. Designed and built by Lavish Sethi.

21
modular pipeline tools in 7 stage groups
98
channel slots managed from one control panel
20+
specialised AI agents (engineering, QA, operations, creators)
10
channels run end-to-end by a single operator agent, with automatic uploads

What it does

Media Factory turns a topic into a finished, publish-ready vertical Short with no manual editing. Each channel slot has its own niche, voice, style and upload pace, and the whole fleet runs from one app.

Lego-style pipeline canvas

Every project is a chain of tool cards. Tools snap together in any order, can be switched on or off, and a needs/gives check blocks combinations that can't work.

Fact-checked by design

Scripts are checked against what is actually said and shown in the source. If no draft passes after a set number of tries, the video is held back instead of published.

Graceful fallbacks

If an optional tool fails, the pipeline uses the previous step's output. If a required step fails, nothing half-finished is ever uploaded.

Channel-safe publishing

Uploads are rate-limited per channel, tagged for traceability, credited correctly and use only licence-safe music and sound effects.

The pipeline

Seven stage groups, 21 tools. Required steps are always on; optional tools are added with “+ Add a tool here” between cards.

1 · Finddiscover & fetch 2 · Momentbest part, trim 3 · Clean upde-caption, upscale 4 · Understandtranscribe, vision 5 · Voicescript, fact check, TTS 6 · Assemblecaptions, music, render 7 · PublishQA gate, schedule fact check fails → redraft (max 3), else hold Orchestrated per channel slot · AI agents monitor, tune and repair each stage
Find & download
  • Topic-based source discovery
  • Reuse & duplicate guard
  • Size-limited fetch
Choose the moment
  • Audio “best moments” model
  • Cut dead air
Clean up
  • Remove burned-in captions
  • AI upscaling
  • Colour boost
Understand
  • Speech-to-text
  • Vision model describes frames
Voice
  • LLM commentary script
  • Automatic claim check
  • Per-channel cloned voice
Sound, text & picture
  • Word-exact captions
  • Animated hook card
  • Licence-safe music & SFX
  • 9:16 render + checks
Publish
  • Fact-checked title
  • Credits & trace tag
  • Paced upload queue
Quality gate: every render is checked for size, audio, length, a clean full decode and required credits before it can enter the upload queue.
Safety gate: per-channel upload pacing, optional human approval, and fallback to the previous step when an optional tool fails.

The multi-agent team

Media Factory is operated like a small company. Each AI agent has one role, clear rules and a reporting line, and a human (the builder) approves anything risky.

Leadership

CEO, COO, CMO and CRO agents set priorities, plan growth and revenue experiments, and coordinate the rest of the team.

CTO

Owns the architecture and roadmap, assigns engineering work and reviews changes before release.

Engineers

Specialists for video, animation, workflow, production and tooling. They build and fix pipeline tools, always backing up before changes.

QA engineer

Tests every change and every new tool against sample renders, and blocks releases that break output quality.

Creator agents

One “YouTuber” agent per channel slot owns its niche, voice, content plan and pipeline settings, plus an operator agent that runs a 10-channel fleet.

Clippers

Cut highlight Shorts from long-form source footage and route them through the right slot's pipeline.

Coach & advisor

Review analytics, compare what's growing with what isn't, and tune titles, thumbnails and descriptions, logging every change.

Distribution

Repurpose finished videos to other short-video platforms and package reusable parts as products.

Workflow examples

Channel names are anonymised, and these runs are illustrative examples of how the tools combine, not logs of specific videos.

Channel A · science facts

  1. Find popular short clips on the topic
  2. Understand the clip: transcript + frame descriptions
  3. Script a 20–40 s explainer; claim check passes on draft 2
  4. Voice in the channel's own voice, word-exact karaoke captions
  5. Render & check, then queue for the next safe upload slot

Channel B · history facts

  1. Best moments picks the most engaging segment
  2. Remove the source's burned-in captions
  3. Animated hook title in the first seconds
  4. Fact-checked title; falls back to a safe title if it fails
  5. Upload with credits and a slot/project trace tag

Channel C · long-form to Shorts

  1. Clipper agent cuts highlights from a long video
  2. Cut dead air to tighten pacing
  3. Upscale low-resolution clips before the 9:16 fit
  4. QA agent spot-checks renders
  5. Coach agent tunes metadata after publish and logs the change

Results

Verified figures come from the system's own configuration and records. Anything marked illustrative is an example, not a measured result.

MeasureValue
Pipeline tools available in the canvas verified21 tools across 7 stage groups
Channel slots managed in one control panel verified98
Specialised AI agents on the team verified20+
Channels run end-to-end by one operator agent with automatic uploads verified10
Script drafts allowed before a video is held back verified3 (configurable 1–5)
Upload pacing per channel verifiedat most 1 every 180 minutes
Manual editing time per Short illustrativefrom roughly an hour by hand to minutes of review

Tech & skills shown

High-level overview of the stack and the engineering skills the project demonstrates.

AI agents & orchestration

Role-based multi-agent team with reporting lines, scheduled routines, human-in-the-loop approvals and change logs.

multi-agentLLM opsHITL

Generative AI

LLM scripting with automated claim checking, vision-language frame description, speech-to-text and voice-cloned text-to-speech.

LLMsVLMsWhisper-class ASRTTS

Video automation

Python + FFmpeg rendering, programmatic motion graphics, AI upscaling, caption removal and audio-based highlight detection.

PythonFFmpegRemotion

Full-stack app

TypeScript web app with a schema-validated, card-based pipeline editor and a local GPU worker that runs the heavy steps.

TypeScriptReactZodSQLCloudflare

QA & reliability

Automated output checks, fallback paths, backups before changes, staged releases and agent-driven regression testing.

testingobservabilityfail-safe design

Responsible automation

Fact-check gate, licence-safe audio only, creator credits, rate limits and AI-content labelling where needed.

complianceplatform safety

The builder

Lavish Sethi

AI automation engineer with a diploma in Computer Programming from Seneca College. Designed Media Factory end to end: product, pipeline architecture, the agent team and its operating rules.

What I'd bring to your team

Shipping practical AI systems: turning a messy creative process into a reliable, observable pipeline, then putting AI agents to work running and improving it safely.

Contact via resume.