Selected work / Client case study
One recording.
Every format, reviewed.
A custom AI-assisted media pipeline that turns a recurring long-form recording into reviewable social videos: editorial suggestions, caption styling, flexible framing, and organized final deliveries in one workspace.
The problem
Many small decisions, scattered.
A recurring recording contains plenty of useful material, but turning it into social content means many small decisions: where to start, what to cut, how to frame it, which captions to use, and which files are actually ready to share.
Those decisions and their outputs needed to stay connected, so the person running it could see which version was selected and what was ready to use.
What we built
A connected review workflow.
A custom application that brings the whole job into one place, with AI assisting the decisions and a person reviewing them:
- Ingest and transcribe the recording, keeping its processing state in one workspace.
- Suggest edits: AI-assisted opening hooks, edited highlights, and a continuous two-minute excerpt.
- Rank music excerpts in a 60–120 second target range for review.
- Review captions: preview and edit timing and wording, with a grouped, highlighted caption style.
- Render every version: landscape and vertical, clean and captioned, with crop, blurred-background, or pan framing.
- Organize delivery: final videos collected into a clearly named folder, with descriptions and supporting files alongside.
One example job
- Source recording
- 1:03:16
- Rendered output versions
- 19
- Speech videos
- 13
- Versions of one 71-second music excerpt
- 6
- Saved caption cues
- 198
Counts describe the captured job: rendered versions, not 19 separate editorial picks. They aren't a guaranteed output or a productivity benchmark.
Feature / Music selection
Find the moment.
The pipeline analyzes music passages and ranks candidate excerpts with music, boundary, continuity, and audio-quality scores. The scores are ranking signals to guide review, not proof that an excerpt is musically complete, so a person listens before choosing.
One selected 71-second excerpt then becomes six delivery versions: three framing options, each clean and captioned, with short audio fades.
See the full music view
Feature / Caption review
Captions, with control.
Caption timing is saved as editable cues, and the style groups three or four words on a blue background with the active word highlighted in yellow. Some timings are estimates, which is why review is built in.
Under the hood
Built to repeat.
A Python application combines a web review interface, saved processing state, AI-assisted analysis, and FFmpeg rendering into a repeatable production workflow. The workspace, files, and rendering run locally; configured AI tasks can call external services.
The result is a clearer path from source material to selected, reviewable deliverables, with editorial choices, output versions, and final files easy to inspect together.
Shown as an anonymized walkthrough of the working application: client names, media, and transcript are concealed, and caption example text is illustrative.
Start with one workflow
WHAT KEEPS
SLOWING YOU DOWN?
Show us the repeated task, the missing information, or the tool that doesn't fit. In a free 20-minute call, we'll decide whether there's a useful project to take on.