Your Tech Club
Signal.Stream logo

Spectre: an AI platform for reality TV that finds the scene

How we built an AI platform for reality TV with Signal.Stream that finds the scenes in thousands of hours of footage. Read how we approached it.

Spectre, the AI platform for reality TV by Signal.Stream and Your Tech Club

Spectre is an AI platform for reality TV. We built it together with Signal.Stream, which makes content for television and streaming. The platform searches thousands of hours of video and audio for the moments that make a season. A confession, a break-up, that one conversation that changes everything.

Moments like these happen once. Then someone has to find them again, and right now that takes a story team days.

I like telling this story, because it shows how I look at AI. First work out where it really saves time. Only then build.

The project in brief

Partner
Signal.Stream, a maker of content for television and streaming.
Sector
television and streaming, unscripted formats.
What we did
AI engineering, platform design and architecture, video and audio processing at scale, UX/UI design, and the design and build of the website.
Technology
a proprietary model (SPECTRE LLM) and Next.js for the website.
Team
Jip van Berkum and Wouter van der Pauw from Signal.Stream, Joey van Vliet and me from Your Tech Club.

What was the challenge in reality TV?

Finding the right scene takes days. Story teams rewatch hours of dailies to find the moments editors build around. General AI models barely help, because they are trained on books, code and the internet. They produce words, but not a moment on the timeline.

A model like that also doesn't know what a confessional sounds like. It loses the thread when six people talk over each other. And it certainly doesn't know the rules of a format.

There is something else too. Unreleased footage is sensitive. You don't put footage from a season that has yet to air into a public AI tool. That applies to more than television, by the way, but here the risk is very tangible.

The page about Spectre's proprietary model, with the heading "An AI built for one industry"

Why did we build a dedicated AI platform for reality TV?

Because adapting a general model solves the wrong problem. Production knowledge without technology would have stayed an idea. Technology without production context would have solved the wrong problem. So with Signal.Stream we built a system for one industry, one that learns the cast, the format and the show bible of a production.

The roles were clear. Signal.Stream brought years of experience in television production, knowledge of editorial workflows and access to real productions. We brought the AI engineering, the design and architecture of the platform, and the technology to process large volumes of material.

That is exactly why it worked. If there is one thing I see in AI projects that stall, it is this: someone starts with the technology instead of the work. Here, we started with a story team's search work.

Does your team also spend hours searching its own material? Then we first look at where AI really saves time, and only build after that.

See how we implement AI
The team behind Spectre: Jip van Berkum and Wouter van der Pauw of Signal.Stream, Yassine Seghrouchni and Joey van Vliet of Your Tech Club

How did we build Spectre?

Together with the people who use it every day. We shaped Spectre from the start with editors, story producers, showrunners and production teams. Their feedback decided not only what the technology could do, but also what it had to do. We designed every feature around a problem they run into daily, with real productions as the test.

The result is a platform that turns raw footage and audio into a story a team can work with. It runs on a proprietary model, built for unscripted television.

Scenes, emotions and conversations

Spectre recognises and organises scenes across all camera angles and shooting days. It flags emotional turning points, conflict and tension. It also maps who said what to whom. Alliances and conflicts become visible, without spreadsheets or memory work.

Spectre's feature overview: scene detection, emotion and tension detection, multi-camera analysis and conversation intelligence

Six people talking over each other

Where a general model gives up when six people talk at once, Spectre separates the voices. It follows speakers across days and cameras. Every moment also comes with the exact second it happened. So an editor lands straight on the right spot in the timeline.

An assistant for the story team

The story team gets daily summaries of what happened and what matters. There is also the AI Spotter, trained on the production's show bible and format. Spectre understands Dutch and the main European languages, because far from every production is in English.

Footage stays with the production

Every production runs in its own isolated environment. Spectre does not use the material to train models for other productions. It encrypts everything in transit, at rest and during processing, and logs every access and action. A production company chooses the data region itself. Spectre can also run in a private cloud, on-premise or on the production's own infrastructure.

Spectre's privacy commitments: no reuse of material, encryption, isolated projects, choice of region and a full audit trail

What does Spectre's design look like?

The design revolves around the timeline. Every moment found points to an exact second, so an editor can carry straight on. Spectre also shows emotion and tension as a graph over time, alongside a timeline that shows per cast member who speaks when. We designed and built the website ourselves too.

That website runs on Next.js. Production teams request access there under NDA, and new places open in rounds. That fits the product: you don't let everyone in at once to footage from a season that has yet to air.

A chart of emotion and tension over time with three peaks, next to a timeline showing which contestant speaks when

What did the AI platform for reality TV deliver?

Five components are now live: scene detection, conversation intelligence, emotion and tension detection, daily reports and the AI Spotter. With these, Spectre takes away the search work, so a story team can spend its time on the story. It doesn't replace an editor's instinct. That was never the intention.

The next step is on the roadmap. Think rough cuts ready for the edit, and direct export to Premiere Pro and DaVinci Resolve. Timelines organised by storyline are coming as well.

If you want to read more about how a system like this fits together, you'll find the basics in what an AI agent is.

Want to see another example from the world of television? For StukTV we built the pin game for Waar de F*ck is STUK?, built for the peak after every call-out.

Want to know where AI really saves time in your business? We work it out in five working days, before you have anything built.

Book your AI audit

Updated

Questions or an intro call

Get in touch with us directly. We'll get back to you as soon as we can

Joey van VlietYassine SeghrouchniThe team at the officeYour message goes straight to our team in Amsterdam.

We'll get in touch as soon as we can.

Frequently asked questions

Who built Spectre?

Spectre is a collaboration between Signal.Stream and Your Tech Club. Signal.Stream brought the knowledge of television production and editorial work. Your Tech Club did the AI engineering, the design and architecture of the platform, the processing of large volumes of video and audio, and the design and build of the website.

What does Spectre do?

Spectre searches the video and audio of a reality production for the moments that matter: emotional turning points, conflicts, conversations and the relationships between cast members, each with the exact second it happened. Scene detection, conversation intelligence, emotion and tension detection, daily reports and the AI Spotter are live.

Why a proprietary model and not a general AI model?

General models are trained on books, code and the internet. They don't recognise a confessional, lose track when several people talk over each other and don't know the rules of a format. Spectre is built for unscripted television and learns the cast, the format and the show bible of a production.

What happens to a production's footage?

Every production runs in its own isolated environment. The material doesn't train models for other productions, is encrypted in transit, at rest and during processing, and every access is logged. The production chooses the data region, and Spectre can also run in a private cloud or on-premise.

Does Spectre work with Premiere Pro or DaVinci Resolve?

Direct export to timelines in Premiere Pro and DaVinci Resolve is on the roadmap for the next version, along with rough cuts ready for the edit.