Back to Projects
privateML/AIbackendml

Omar M Studio — Multi-Agent Video Production Crew

A local AI production crew. Research, writer, director and analyst agents turn a topic into a fact-checked script, a take report, an auto-cut, and a captioned, mixed final video — one video in about 15 minutes.

2026-09 — Present

Key Highlights

  • A crew of research, writer, director and analyst agents takes a topic to a finished 9:16 short
  • Writer runs headless Claude Code under a tool allowlist and lints its script until PASS
  • Director: on-device Whisper (MLX) reviews every take in ~30 seconds
  • Auto-cut removes flubs only when the transcript confirms them
  • 19K lines of Python, 454 tests, 69 logged architecture decisions

Overview

Omar M Studio is the production system behind my short-form channel on how engineers actually work with AI. AI finds the topic, drafts and edits; I record every take.

The crew

  • Research scans vendor feeds, arXiv and policy-page diffs, and scores each lead by how little it's been covered yet.
  • Writer runs headless Claude Code under a fixed tool allowlist; it drafts a sourced script and loops a linter until PASS — no commit rights, redacted event log.
  • Director runs Whisper large-v3 on-device (MLX) to check pace, energy, fillers and drift from the script, then says keep or retake — in about 30 seconds.
  • Analyst OCRs retention curves and ties each drop to the script line playing at that second.

    • Finish

    The auto-cut removes flubs only when the transcript confirms them. Finish burns captions with the script's spelling, places sound effects on words, ducks music and runs two-pass loudness normalization.

    By the numbers

  • 107 commits in 30 days, built solo
  • 19K lines of Python, 454 tests
  • 69 logged architecture decisions