DCT

Audience intelligence that knows the person, not the pattern

We build the ranking that decides what each viewer or shopper sees next, the assistants that answer them, and the campaign tooling marketers use without an engineer.

Talk to a DCT AI EngineerSee how a recommendation is served ↓

  • 85%

    faster campaign production

  • 90%

    faster publishing

  • 50,000+

    subscribers in quarter one

  • Day one

    personalization live at launch

How a recommendation is served

Four decisions, one request

Personalization looks like one decision and is really four: which items are even candidates, which of those this person is allowed and likely to want, how to order them, and how much room to leave for something they have not seen before. Get the fourth one wrong and the catalogue quietly shrinks to whatever already worked.

Swipe to see more →
LIVE SIGNALthis sessionwatch historycatalogue signalcampaign & lifecycleCandidates1,842 retrievedEligibility311 remainRanktop 20Explorereserved slotsSurfaceapp · player · inboxBRAND · RIGHTS · SAFETYenforced here, not reviewed afterwhat they did next becomes tomorrow's signalHOW IT IS JUDGEDHOLDOUTA slice of traffic never sees the change. Offline scores decide what is worth testing;the holdout decides what ships - measured on what people came back for, not once.
Active stage

Candidates

Retrieval, not scoring. A few thousand plausible items are pulled from a catalogue of hundreds of thousands, using cheap signals - recent behaviour, similarity, what is live, what is trending in this market. Anything that does not make it here can never be recommended, which makes this the least glamorous and most consequential step.

01 - Signal

Behaviour beats attributes, and freshness sets the ceiling

A segment built at registration describes who someone was. What they did in the last four minutes describes who they are now, and it is the stronger predictor by a wide margin. So the first work is rarely modelling - it is getting behaviour off the surfaces where it happens and into a stream something can act on inside a request.

  • Session signal, not just profile
  • One stream across app, player and campaign
  • Latency budget agreed before modelling
02 - Cold start

The hard part is what nobody has watched yet

Ranking things with a history is comparatively easy. A new title, a new product, a new customer has no history, and a system that only ranks on behaviour will bury all three forever. We rank new items on what they are - the content itself, its metadata, its similarity to things that worked - and keep reserved slots so new material gets seen at all.

  • New items ranked on content, not history
  • Reserved exploration slots per surface
  • New customers served from the first session
03 - Measurement

NBA decides what to try. Live traffic decides what ships.

DCT brings together product thinking, engineering expertise and deep technology experience to build, modernize and scale digital products that create lasting business value.

  • Holdout on every change
  • Return and completion over click
  • Watch for the catalogue narrowing
The same catalogue · A different answer per person

Typical Personalization vs. In-the-Moment Personalization

  • Segments defined by hand and refreshed monthly
    Segments computed from what someone is doing now
  • Recommendations from last night's refresh
    Ranked at request time, on the session so far
  • New titles wait until enough people have found them
    New items ranked on what they are, from day one
  • Re-engagement goes out on a calendar
    Triggered by the behaviour that means something
  • Marketing waits on engineering to build the asset
    Marketers produce and publish, with approvals intact
  • Brand and rights checked in review, after the fact
    Enforced before anything can reach a customer
  • Success reported as a click-through rate in a deck
    Measured against a holdout on returning customers
  • The catalogue quietly narrows to what already worked
    Exploration is budgeted, so discovery keeps working
Claude Preferred Partner

Claude Partner Network

Three places, not everywhere. Reading a catalogue item well enough to rank it before anyone has watched it. Producing the customer-facing asset a marketer described. Answering a customer directly, against their real account and content. The scoring path stays fast and measurable. DCT builds AI capabilities on an enterprise-grade model and deployment architecture.

  • Understanding the catalogue

    New items described richly enough to be ranked before behaviour exists - the cold-start problem, answered with content.

  • Producing the asset

    A brief or a reference image becomes a production-ready, brand-compliant campaign the marketer can edit in plain language.

  • Answering the customer

    Conversational assistants grounded in real account and content context, handing over to a person with the history attached.

  • Cost per thousand requests

    Serving cost is a design constraint at recommendation volume, measured from the first week.

Customer stories

Where the production cycle was longer than the moment

  • Sports streaming · Digital media

    Campaign brief to published email in under an hour

    Marketing tempo was set by the sporting calendar, which does not wait. Every campaign needed marketing, design and development to coordinate - building responsive email, holding brand consistency, managing personalization, testing across devices, publishing. Each step had its own queue, and the whole cycle was longer than the moment it was meant to respond to.

    DCT built an agentic campaign tool inside the client's own tooling. A marketer describes the campaign in plain language or drops in a reference screenshot, iterates in chat against a live preview, and approves. The platform versions it and publishes it - with approvals, ownership and audit unchanged.

    • 85%

      faster email production

    • 90%

      faster publishing

    • 80%

      less personalization setup

    Engineering left the critical path entirely. Speed changed what the team could attempt - campaigns now respond to a live result rather than anticipating one, and iteration happens because iteration became cheap.

  • Media & entertainment · Short-form video

    A short-form service launched with personalization already in it

    Audience behaviour had shifted to mobile-first, bite-sized viewing and the window was open now. Short-form is unforgiving: sessions run in minutes against a catalogue of thousands of clips, so discovery is the product rather than a feature on top of it. Building the platform from scratch would have put the launch a long way out.

    DCT customized and launched its existing short-form product for Android and iOS, with behavioural personalization, continuous segmentation and automated re-engagement live at launch - not deferred to a second phase.

    • 50,000+

      subscribers in quarter one

    • Two

      mobile platforms at launch

    • Day one

      personalization live

    The launch took a customization cycle, not a build cycle. And the intelligence arrived with it - which is the part that usually gets deferred to phase two and never happens.

Common questions

What product and growth leaders ask first

DCT Pulse is delivered by DCT AI engineers working inside your team, on your surfaces and your data. These are the questions that come up before anyone signs anything.

What's the ROI of adding AI personalization to an existing recommendation engine?

The gains from AI personalization usually don't come from replacing the ranking model, they come from what surrounds it. What makes it into the candidate set, how fast a session's signal reaches the request, how items with no history get handled, and whether results are measured against a holdout instead of last month. DCT Pulse assesses all of this first, so you see the actual ceiling before any rebuild starts.

How does AI personalization handle new products or new customers with no data?

AI personalization handles the cold-start problem by ranking on content and context instead of waiting for behavioural history. New items rank on what they are, metadata, similarity to what's performed, so they're placed sensibly from day one. New customers get served on session signal instead of waiting for a profile to build. DCT Pulse also reserves a share of slots for items it's still uncertain about, validated against real traffic before they're trusted.

How do you measure whether AI personalization is actually working?

AI personalization is measured against a holdout, not a dashboard. A slice of traffic never sees the change, so the comparison is real instead of a before-and-after on a moving baseline. The measures that matter are lower bounce, higher completion and return rate, and whether the recommended next action actually gets taken. Click-through isn't the headline metric, it's the easiest one to win badly.

Does AI personalization create a filter bubble over time?

Yes, if nobody's watching for it. A system trained only on clicks keeps showing what got clicked, and the effective catalogue quietly shrinks, even as engagement numbers look fine. DCT Pulse tracks catalogue coverage as a first-class measure alongside engagement, and reserves exploration slots specifically to counter this.

Who controls what DCT Pulse can show a customer?

You do, and it's enforced before anything reaches a surface, not caught in review afterward. Rights, territory rules, brand standards, and safety constraints gate the surface directly, so anything not clearable in a market can't be ranked into it. Where content is generated rather than selected, a campaign, an assistant reply, every asset stays versioned, attributable, and reversible.

How fast does DCT Pulse run, and what does it cost to serve recommendations?

A recommendation that arrives after the page renders isn't a recommendation. The latency budget is set before modelling starts, so heavier work happens ahead of time and gets looked up, not computed live. Cost per thousand requests is treated the same way, a constraint from week one, not something discovered once volume arrives.

What do you need from us to get started with DCT Pulse?

Access to the behavioural signal your surfaces already produce, and agreement on which surface goes first. We baseline current performance, deploy on that one high-traffic surface, and measure against it before extending to adjacent surfaces and lifecycle stages, with governance defined before scale, not after.

Start on one surface, against a holdout

0/255