Intelligence · Circo

VideoSense: an AI video intelligence engine

A pipeline that watches a video the way an editor would: scenes, speech, motion and narrative in; teasers, platform-ready presets, thumbnail crops and monetisation signals out.

Architect · Full-Stack Engineer

The problem

Creators upload a finished video and then face hours of packaging work: finding the best moments, cutting a teaser, writing titles for each platform, choosing a thumbnail. Circo's studio, broadcast, series and reach tools all needed the same understanding of a video to automate that work.

What it does

VideoSense is the intelligence layer embedded in those tools. It analyses audio, scenes, speech, motion and narrative, and produces outputs a creator can use straight away: teaser clips, platform-specific upload presets, title and description variants, normalised hashtags, thumbnail crop recommendations and monetisation signals.

How it works

Facts first. The first stage is deterministic: duration, scene changes, silence, speech, topics, audio tone and a summary, all measured rather than guessed.
Then judgement. AI stages build on those facts: narrative beats, hook scoring, quality assessment, clip ranking and a safety check.
Then packaging. A teaser engine selects and packages clips per platform, and a final stage assembles titles, crops and upload presets.
Additive by design. Twenty-two steps in four layers, each layer extending the one before. There is no version branching anywhere: every video runs the full chain, and later layers only ever add fields.
Fails soft at the edges. If a packaging step fails, the rest still run and earlier results are never lost. Only a failure in the core analysis fails the job.
Testable AI. Every model call goes through one interface, so the test suite runs against a deterministic stand-in, with more than four hundred tests and no network calls.

Why it matters

Packaging is where creators lose time and reach. Turning one upload into everything each platform needs, automatically, gives that time back, and the additive design means every new capability compounds on the last instead of forking the pipeline.

System

Video moderation that fails closed

A media admission pipeline that probes cheaply, gates on spec, classifies with a model, and always returns one predictable result, rejecting what it cannot inspect.

ModerationContractsReliability

System

One upload, many episodes

A long upload can be cut into time-based episodes and announced to the platform as a series, without the creator editing anything.

EpisodesMedia processing

System

Adaptive streaming and the transcoding ladder

Every upload becomes a ladder of renditions, packaged once for HLS and DASH and served from the edge, so every phone gets the best stream its connection can hold. Live in production.

TranscodingABRCMAF

System

Personalised feeds on a recommendation engine

Every interaction batched into a recommendation engine, with visibility rules and session continuation so the feed stays safe and attributable.

RecommendationsFeeds

VideoSense: an AI video intelligence engine

The problem

What it does

How it works

Why it matters

Related systems

Video moderation that fails closed

One upload, many episodes

Adaptive streaming and the transcoding ladder

Personalised feeds on a recommendation engine

The problem

What it does

How it works

Why it matters