Recommendation systems
Showing everyone the same thing is a decision. It's just one you made by not making it.
The cost of the default
If your product has more items than a person will look through — products, articles, courses, listings, tracks — then most of your catalogue is invisible to most of your users.
They see the homepage, the first page of results, whatever's featured. Everything else exists but is functionally absent, which means inventory you paid for, content you commissioned, and listings someone is paying you to display all sit unseen.
Recommendation is how you close the gap between what you have and what each person sees.
What we build
- Product recommendations
- Related items, complements, and personalised merchandising for e-commerce.
- Content recommendation
- Next article, next video, next course. Keeping people engaged past the thing they arrived for.
- Search ranking
- Personalising result order, which is often higher value than a recommendation carousel and considerably less visible.
- Matching systems
- Two-sided marketplaces: candidates to roles, providers to customers, supply to demand.
- Email and notification targeting
- Deciding what to send whom, and critically when not to send at all.
The problems that define the work
- Cold start
- New users have no history, and new items have no interactions. Both need a sensible answer from day one — content-based similarity, popularity fallbacks, quick preference capture — or your newest inventory never surfaces.
- The feedback loop
- A recommender trained on its own recommendations narrows relentlessly. It shows popular items, which become more popular, which get shown more. Left alone it converges on a small slice of your catalogue and calls it success. Exploration has to be designed in deliberately.
- Diversity and filter bubbles
- The most accurate recommendation is often the most boring one. Five variations of the thing they just bought is technically well predicted and commercially useless.
- Business rules alongside the model
- Margin, stock, contractual placement, seasonal priorities. The model produces relevance; your business decides what to do with it. Both matter and they have to be composable.
How we measure it
Offline metrics tell you a model learned something. They don't tell you it makes money.
So we build for A/B testing from the start, against the metric you actually care about — revenue per session, retention, engaged time, application completion. If the new system doesn't beat the current one on that number, it doesn't ship, regardless of how good its offline scores are.
Before you ask.
Less than people assume to start. Content-based approaches work from item attributes alone and improve as behavioural data accumulates. We'll design for whichever stage you're at.
Machine Learning Models
A model that scores 94% in a notebook and a model that earns its keep are different achievements. The distance between them is where this work lives.
Data & CloudData Engineering
Nobody has ever asked us for data engineering. They ask for the dashboard, the model, or the AI feature — and then we find out why it isn't working.
Data & CloudPredictive Analytics
A forecast nobody acts on is a hobby. The prediction is the easy half; the hard half is what happens next.
Tell us what you're trying to build.
A 30-minute call, no charge and no pitch deck. Describe the problem and we'll tell you how we'd approach it, roughly what it costs, and whether we're the right team for it. If we're not, we'll say so.
30 minutes · No charge · No deck