User-based and item-based collaborative filtering trained on your interaction data -- purchase history, click streams, view events, ratings, and engagement signals. Matrix factorisation approaches (ALS, SVD) for large-scale user-item interaction datasets. Real-time user similarity computation for personalised recommendations. Cold-start handling for new users with content-based fallbacks.

Item similarity models built from catalogue attributes -- product categories, tags, descriptions, price ranges, and custom metadata. User preference profiles built from interaction history. Hybrid content-item representations that combine structured attributes with text embeddings from product descriptions. Effective for catalogues with rich metadata and for new-item cold-start scenarios.

Recommendation systems that use large language models to understand item descriptions, user queries, and preference signals in natural language. Semantic similarity between user intent and catalogue items. Recommendation explanations in natural language ("Recommended because you bought X"). Effective for conversational recommendation interfaces and for catalogues where text descriptions carry the primary signal.

Low-latency recommendation APIs that serve personalised recommendations in real time -- typically under 100ms for homepage, product detail page, and cart recommendations. Precomputed recommendation caches for high-traffic surfaces, real-time user event processing for recency weighting, and feature stores that make user context available to the recommendation model without repeated computation.

Batch recommendation pipelines for personalised email and push notification content -- product recommendations, content suggestions, and re-engagement recommendations based on user history and current context. Scheduled generation of personalised content for sending time optimisation. Integration with email platforms (Klaviyo, Mailchimp, Iterable) and push notification services (OneSignal, Firebase Cloud Messaging).

Batch jobs run on a configurable cadence (nightly for daily sends, hourly for triggered workflows) using precomputed recommendation vectors stored in Redis or a feature store such as Feast. Popularity-based fallback recommendations handle cold-start users who have no interaction history, pulling from trending items within the user's most-visited categories. Segment-level personalisation groups users by behavioural cohort when individual histories are too sparse for reliable collaborative filtering. Re-engagement sequences use recency-weighted item embeddings to surface products the user showed intent on but did not purchase, rather than generic bestsellers.

Experimentation infrastructure for recommendation systems: user assignment to control and treatment groups, metric tracking for business impact (CTR, conversion, revenue per user, engagement), statistical significance testing, and reporting dashboards. A/B testing that tells you whether your recommendations are actually driving the outcomes you care about, not just whether they look different.

Holdout groups are assigned at the user level and hashed consistently so users stay in the same bucket across sessions. Online metrics tracked per experiment include click-through rate at position k, add-to-cart rate, conversion rate, revenue per impression, and session depth. Offline evaluation during development uses precision@10 and NDCG (normalised discounted cumulative gain) measured on a time-based held-out split -- not a random split, which leaks future signal into training. Bandit algorithms (epsilon-greedy, Thompson sampling) are available for recommendation contexts where a hard A/B split wastes too much opportunity cost on clearly inferior variants. Experiment duration is calculated from expected traffic volume and minimum detectable effect before any test is launched, so you know upfront whether the test will reach statistical significance before your product cycle closes.

Before building, we assess your data -- interaction volume, catalogue size, metadata quality, and cold-start severity. The assessment determines which recommendation approach will work for your specific data state. We don't recommend collaborative filtering if you don't have sufficient interaction data, or content-based filtering if your item metadata is sparse. Honest assessment before any development commitment.

Every recommendation model is evaluated on historical data before deployment: precision and recall at K, NDCG, coverage, and novelty metrics measured on a held-out test set. Offline evaluation catches approaches that look good on average but fail on specific user segments or catalogue sections. We establish minimum performance thresholds before the model goes to production.

Production recommendation systems improve over time through experimentation. We build the A/B testing infrastructure so your team can run controlled experiments on recommendation changes and measure the actual business impact. Recommendation quality is tracked as a product metric, not a one-time engineering deliverable.

Recommendation APIs integrated into your product -- e-commerce platform, mobile app, content management system, or custom application. Event tracking for interaction data collection (views, clicks, purchases, ratings) that feeds back into model retraining. Data pipeline from your product database to the recommendation model. The full integration, not just a model.