Skip to content
dhii.ai

Model Training & ML

Recommendation Systems

Show each user the products, content or matches most likely to be useful to them — and prove the lift with a proper experiment.

Baseline recommender
2–3 weeks
Learned ranking
4–8 weeks
A/B test period
2–4 weeks

Relevance is a product feature

Whether you run a store, a content platform, a learning product or a marketplace, users judge you by how quickly they find something worth their time. Generic “bestsellers” and “new arrivals” lists help a little; a recommendation system that understands each user’s context helps a lot more. In our own product, OptyCV, matching candidates with relevant openings is the core of the value — the same family of techniques applies to products, articles and courses.

A pragmatic architecture

Modern recommenders work in two stages. Candidate generation quickly narrows a large catalogue to a few hundred plausible items using collaborative filtering (“people like you engaged with”) and embeddings (“items similar to what you engaged with”). Ranking then orders those candidates with a model that considers the user, the item and the moment — device, time, recent activity — before business rules apply stock, margin, diversity or compliance constraints.

We always start with a simple baseline and only add complexity when it wins offline and in an online test.

Cold start and fairness

New users and new items have no history. We handle them with content-based signals, onboarding preferences and sensible popularity fallbacks. We also measure what accuracy alone hides — diversity, catalogue coverage and whether certain sellers or content creators are systematically buried — and tune for a healthy balance.

Proving the lift

Offline metrics are a guide; the truth is an experiment. We set up A/B testing with a clear primary metric and guardrail metrics, run it long enough to be meaningful, and report the result plainly, including if the new system didn’t win.

Deliverables

What you get

  • 01

    Signal audit

    Review of interaction data — views, clicks, purchases, ratings — and catalogue metadata.

  • 02

    Candidate generation

    Fast retrieval of relevant items using collaborative filtering and embeddings.

  • 03

    Ranking model

    A learned ranker using user, item and context features, with business rules applied.

  • 04

    Cold-start strategy

    Content-based and popularity fallbacks for new users and new items.

  • 05

    Recommendation API

    A low-latency service your site, app or emails can call.

  • 06

    Experiment framework

    A/B testing and metrics so you can measure real impact on conversion or engagement.

Process

How a recommendation systems project runs

  1. Step 1

    Define success

    Choose the metric that matters — revenue, engagement, retention or match quality.

  2. Step 2

    Baseline

    Start with simple strategies like popularity and co-occurrence.

  3. Step 3

    Model

    Add embeddings and learned ranking where they beat the baseline offline.

  4. Step 4

    Experiment

    A/B test online against the current experience.

  5. Step 5

    Scale

    Roll out the winner and keep improving with fresh data.

Engagement models

Ways to work together

Every engagement starts with a fixed, written quote. See engagement models & pricing.

FAQ

Recommendation Systems: common questions

How much data does a recommender need?

Collaborative methods need enough user-item interactions to find patterns; content-based and embedding methods work from item descriptions even with little history. Most real systems blend both.

Can LLMs improve recommendations?

Yes, in specific ways: generating better item descriptions and embeddings, explaining recommendations, and re-ranking a short candidate list using richer understanding. Using an LLM to rank an entire catalogue per request is usually too slow and expensive.

How do you avoid recommending the same popular items to everyone?

We measure diversity, novelty and catalogue coverage alongside accuracy, and tune or post-process rankings to balance them.

How will we know it's working?

Through an A/B test against your current experience, measured on a business metric you choose, not just offline accuracy.

Related services

Start a project

Have a workflow that should run itself?

Tell us what you want automated or built. You get a written assessment of feasibility, risks and a fixed quote — usually within two business days.