Skip to content

Own product · AI video pipeline

ContentVEO: Short Clips From Long Videos

ContentVEO is our own software for turning long videos with English speech, such as podcasts, interviews, webinars and lessons, into captioned vertical clips. AI picks the moments, face tracking reframes them to 9:16, a quality check rejects weak clips, and the rest are scheduled to social platforms. It began as an automation for a podcast client and became a multi-tenant product in September 2026.

The problem

What needed solving

A long podcast or webinar holds several short clips worth posting, but cutting them by hand means watching the whole recording, finding the moments, reframing a wide shot into a vertical one, adding captions and a cover, and then uploading the result to each platform separately. For a weekly show that is a lot of repetitive editing.

ContentVEO started as that job for one podcast client: take each new episode and produce ready-to-post clips automatically. Once the pipeline was reliable, we rebuilt it as a product so other creators could use it with their own accounts.

Constraints

What the build had to respect

  • Compute costs must stay bounded

    Video processing is expensive, so each video gets a limited number of attempts and a time limit, users can stop a job, and the cloud project has a budget cap that pauses processing automatically if spending runs past it.

  • Each account sees only its own data

    Ownership always comes from the verified login token, never from what the browser sends. Every change to database security policies is snapshotted before and after, so it can be rolled back precisely.

  • Social accounts need care

    OAuth tokens for connected social accounts are stored encrypted, and each platform's API requires its own app review before publishing is allowed.

  • Media is not kept forever

    Uploaded source videos are deleted after a few days and finished clips after a period that depends on the plan, as set out in the retention policy.

What we built

The build, part by part

A React dashboard on Supabase, with the heavy video work on a Google Cloud Run worker and scheduled jobs in GitHub Actions.

Ingest and transcribe

  • Videos from Google Drive, a direct upload or a link
  • Transcription through Groq, falling back to a local whisper.cpp model

Pick the moments

  • Audio analysis to find lively passages
  • AI clip selection with a fallback chain across Gemini, OpenAI, Anthropic and OpenRouter

Make the clips

  • Face-tracked reframing to vertical 9:16 with MediaPipe
  • Rendering with ffmpeg, burned-in captions and generated covers
  • An AI vision check that rejects clips that fail quality rules

Review and publish

  • A clip review screen and a publishing calendar
  • Publishers for YouTube, TikTok, Instagram, Facebook, LinkedIn and X
  • Brand kit, Slack notifications, usage and cost views, and Stripe billing with a minutes ledger

One known limit is visible in the test set: when a speaker leans out of frame, face tracking can cut part of the face off. The quality check catches those clips and drops them instead of publishing them, which is why a source video sometimes yields fewer clips than it could.

From a long video to scheduled clips
  1. Ingest

    Drive, upload or link

  2. Transcribe

    Groq or local whisper.cpp

  3. Pick moments

    Audio analysis and AI

  4. Reframe

    Face tracking to 9:16

  5. Render

    Captions and covers

  6. Quality check

    Weak clips dropped

  7. Review and schedule

    To connected platforms

Stack and integrations

App
  • React
  • Vite
  • Tailwind
  • Supabase
Processing
  • Google Cloud Run
  • ffmpeg
  • MediaPipe
  • whisper.cpp
  • yt-dlp
AI
  • Groq
  • Gemini
  • OpenAI
  • Anthropic
  • OpenRouter
Operations
  • GitHub Actions
  • Stripe
  • Netlify
  • Hostinger

Handover

How we test and run it

As with Invollo, there is no client handover, so the work goes into testing and security.

  • Test suites run in continuous integration, plus a "golden set" of reference videos re-processed after changes
  • A documented end-to-end audit in October 2026 covering the customer journey, billing, the worker and production smoke tests
  • A security release that fixed the findings of a review, including server-side request forgery, OAuth state checks, option injection in the video downloader and queue abuse
  • Launch checklists for billing, platform reviews and cost controls

Status and link

Where it stands

Live at contentveo.com. Google approved the YouTube API review; as of November 2026 the TikTok and Meta app reviews, which cover TikTok, Instagram and Facebook publishing, were still pending.

Visit contentveo.com (opens in a new tab)

What we leave out

What we do not claim

No user, revenue or engagement figures: the product is new and we have not audited any numbers we could publish. The clip counts and scores inside the app describe a video, not a business result.