Skip to content
Back to work

AI & Automation · Retail & Commerce

Recommendation Engine

For a retailer with a large catalog, a recommendation engine that helps customers find relevant products, gives new listings a route to visibility and keeps merchandisers in control.

Built on
Python, TensorFlow, Redis, FastAPI
Industry
Retail & Commerce
Project type
AI & Automation
Primary service
AI & Automation
Scope
Model design, serving architecture, business rules layer, experimentation design.

Repository not shown as per company policy

01Challenge

A large catalog is only an asset if people can find their way through it.

  • Popularity bias

    Recommendations based on co-purchase alone kept promoting the same small set of products.

  • Cold start

    New listings had no history, so the existing approach never recommended them.

  • No commercial control

    The merchandising team could not steer what appeared, even for stock that needed to move.

02Approach

Approach

Blend behavioral signals with product attributes so the system has something sensible to say about items with little history, and keep the ranking explainable enough for merchandisers to trust and override.

03Outcome

Recommendations that help discovery, not just repeat the bestsellers.

Generic “customers also bought” strips kept recommending what was already popular. New products and slower lines never got seen.

  • Behavior plus attributes

    Blending what customers do with what products are gives the model something sensible to say about items with little history.

  • Merchandisers stay in charge

    A rules layer above the model lets the team pin, exclude and boost products without retraining anything.

  • Evidence before rollout

    Ranking strategies are compared in controlled experiments rather than swapped on instinct.

What changed

  • Recommendations draw on product attributes as well as behavior, so new and low-traffic products can still be recommended.
  • Merchandisers pin, exclude and boost products through a rules layer that sits above the model.
  • Ranking strategies are compared through controlled experiments before they are rolled out.

How it works

Candidates, ranking and rules as separate stages.

overridessplit trafficBehaviorViews, carts, purchasesCatalogAttributes, categoriesCandidate generationCached candidate sets,refreshed on a scheduleMerchandisersPin, exclude and boostwithout a releaseLow-latency APIFastAPI with Redis cacheExperimentsStrategy comparisonRankingCandidates, ranking and rules as separate stagesHybrid rankerCollaborative and content-basedBusiness rulesPin, exclude, boost: merchandiser overrides

One run, step by step.

01 / 05

Inside the system

What it does.

  • Hybrid behavioral and attribute-based ranking
  • Cold-start handling for new and low-traffic products
  • Low-latency serving with cached candidate generation
  • Merchandiser override and business rules layer
  • Experiment framework for comparing ranking strategies

Built with

What it runs on.

Modeling
PythonTensorFlow
Serving
FastAPIRedis
Control
Merchandiser rules layerExperiment framework

Customers find more of the catalog, and the team still decides what matters.

More work

Related projects.

AI & AutomationHealthcare

Autonomous AI Log Monitoring & Observability Platform

For a healthcare media and clinician engagement platform, a five-agent system that reads a production error, writes the fix and opens a reviewed pull request, with an engineer still deciding what ships.

AI & AutomationHealthcare

Drug Competitor Identification

A brand-intelligence tool that asks a language model who a drug competes with, then checks the answer against regulatory reference data before anyone is asked to trust it.

AI & AutomationHealthcare

AI Agents Platform

Seven agents that turn one upload into a recorded, print-ready batch of personalized posters, with a reviewer approving anything that carries commercial risk.

Discuss a similar project.

If something here is close to what you need, tell us about your situation and we will walk you through how we would approach it.

Intelligence → Innovation → Automation → Growth