Media Automation
Pipelines.
We build media pipelines that run in production. OutCaps, our localization engine, takes video ads into new languages and keeps the subtitles locked to the speaker — nothing to re-time, no line breaks to fix by hand.
We don’t answer physics questions with a language model.
When a pipeline misbehaves, the usual fix today is more tokens — another prompt layer, another agent reviewing the last one. Building OutCaps took us the other way: layout, timing and fit are arithmetic, so they are solved in code, and the model is spent only where judgment is genuinely required. The result is a pipeline that behaves the same way on every run — which is what lets a team depend on it.
Don't ask a model to do arithmetic
Line width, reading speed, safe-zone fit — these have exact answers. A model can only estimate them: it doesn't know how wide a word renders in a given font unless you hand it the metrics, and at that point you've already built the deterministic tool. In OutCaps that math is plain code — same input, same output, every run.
The model translates. A compiler makes it fit.
Meaning, tone and idiom are the model's job. Everything physical — where a line breaks, how long it holds, how it scales — is a separate deterministic pass that runs afterwards. That split is why OutCaps is two-stage: asking one prompt to do both jobs is how a caption ends up grammatically perfect and unreadable on screen.
Nothing is a black box you can't edit
Every stage of the pipeline writes an artifact a person can read, diff and correct — the text layer is never skipped. That matters on the day a brand name comes out wrong: the fix is editing one line and re-exporting, not re-generating everything and reviewing it again from the start.
People get the decisions, not the cleanup
The pipeline surfaces the handful of moments where judgment actually matters — a line that can't fit at readable speed, a phrase with two defensible readings — instead of returning a wall of output to patch. We learned this from real review sessions: when the reviewer is dragging boxes and re-timing lines, the work wasn't automated. It was moved.
Notes from the pipeline.
Occasional engineering notes on automating real-world work — where deterministic code beats another agent loop, what a black-box model costs you the moment something needs fixing, and how agentic workflows behave outside the demo, against constraints that don’t negotiate. Localization is where we started, not the subject.
By subscribing, you agree to our . You can unsubscribe at any time.