Industries · Retail and marketplaces

Better product discovery and retail operations, measured against what you have today

If your retail business or marketplace needs better search, recommendations, catalogue data or support automation, I start by ranking the candidate problems by measurable effect and data readiness, then build one pilot evaluated against your current baseline. Often the right answer is conventional retrieval or ranking rather than generative AI. You keep the pilot and a rollout plan your team can own.

This is a good fit if…

  • Shoppers cannot find products you stock: zero-result searches, failed synonyms and attribute queries, or long-tail items that never surface.
  • Recommendations are hand-written rules or a vendor black box, and nobody can say whether they help.
  • Your catalogue has inconsistent attributes, duplicates or thin descriptions from many sellers or suppliers.
  • Support volume is dominated by order, returns and product questions that your existing content could already answer.

Look elsewhere if…

  • You want prices set automatically without human approval. I do not take on automated pricing as a first pilot.
  • You need a commerce re-platform. This work sits on top of your existing platform and search stack.
  • You already run a retrieval-based assistant that gives wrong answers. Use RAG quality rescue instead.

What you get

Retail-specific opportunity brief, evaluated retrieval or recommendation pilot and rollout plan

  • A ranked opportunity brief covering search, recommendations, catalogue enrichment and support, each with its measure and data dependency.
  • One pilot evaluated offline against a labelled set and, where traffic allows, in a controlled online test.
  • A reasoned choice between conventional ML, retrieval and a generative model for each problem.
  • Event-tracking and catalogue-data gaps written up with owners.
  • A rollout plan with guardrails, fallback behaviour and a named owner.

How it runs

  1. 01

    Opportunity brief

    I review search logs, catalogue data, behavioural events and support contacts, then rank candidate problems by measurable effect, data readiness and commercial risk.

  2. 02

    Baseline and measure

    Before building, we fix the measure: offline relevance on a labelled query set, plus the online metric and test design you will use, such as search exits or contact rate.

  3. 03

    Pilot

    I build the chosen change in your stack behind a flag: hybrid search, re-ranking, attribute extraction, recommendation candidates or a support answer flow.

  4. 04

    Decision and rollout plan

    Results against the baseline, a go, revise or stop recommendation, rollout guardrails and handover to your product and engineering owners.

What needs to be in place

  • Access to search logs, catalogue data and behavioural events, or a representative export.
  • A product owner who can decide what 'better' means and accept trade-offs between relevance, margin and stock.
  • An engineering owner and a way to ship behind a flag or to a test segment.
  • Merchandising or category experts for a few hours a week to label relevance.

Not included

  • Promised conversion, revenue or engagement figures.
  • Automated price changes without human approval.
  • Replacing your commerce platform or renegotiating your search vendor contract.
  • Ongoing merchandising operations or catalogue data entry.

The situation

Retail and marketplace teams rarely lack AI ideas. The backlog already holds semantic search, personalised recommendations, generated product descriptions, a support assistant and dynamic pricing. What is missing is a way to choose between them, and a measure that will tell you, after the work, whether it helped.

In retail that measure matters more than in most sectors, because mistakes are visible and commercial. A search change that surfaces out-of-stock items, a recommendation carousel that pushes low-margin products, or a support answer that misstates the returns policy shows up in revenue and contact volume within days. The work therefore starts with the baseline, not the model.

What the work involves

Search and discovery. Most failures come from vocabulary and structure, not from the lack of a large model: shoppers use words your catalogue does not, attribute queries (“waterproof, size 9, under a budget”) are not parsed into filters, and lexical search misses close matches. Hybrid retrieval, combining lexical and vector search with metadata filters and a re-ranker, usually beats either alone. I build a labelled query set from your logs with your category experts, so every change is measured on the queries your shoppers actually type.

Recommendations. A production recommender has three layers: candidate generation (fast and recall-oriented), ranking (slower and precision-oriented) and re-ranking for business rules such as stock, diversity, freshness and margin. The largest gains are often in the data underneath: clean event tracking, de-duplicated identities and consistent product attributes. I built a Word2Vec-based recommendation system for Lenskart and designed the content recommendation engine and data pipeline for Momspresso; both depended on getting that foundation right.

Catalogue quality. Marketplaces and multi-supplier retailers inherit messy listings. Language models are useful here for extracting structured attributes from supplier text, flagging duplicates and drafting descriptions for a merchandiser to approve, with every generated field traceable to its source.

Support. Order-status, returns and product questions can often be answered from existing policy and catalogue content. I have integrated AI assistance into the trade chat of a peer-to-peer marketplace; the lesson carried over is that escalation to a person, and keeping policy content current, matter more than the model.

The signature deliverable, illustrated

The first output is an opportunity brief. Illustrative example, not a client result:

CandidateMeasureData readinessRiskRecommended approach
Zero-result searchesShare of searches with no results; search exit rateLogs available; synonyms unmanagedLowHybrid retrieval plus synonym mining
”You may also like” carouselAdd-to-basket from carousel in a controlled testEvents incomplete on mobile appMediumFix tracking first, then candidate model
Supplier attribute gapsShare of listings with complete key attributesSupplier feeds vary by formatLowExtraction with merchandiser approval
Returns-policy questionsContact rate on returns topicsPolicy pages currentMediumAnswer flow with escalation

The brief is followed by one evaluated pilot and a rollout plan stating guardrails (stock and margin rules, fallback to current behaviour), monitoring and owner.

How acceptance is judged

The measure, the baseline and the decision rule are written into the scope before the build. Your product owner accepts the pilot on those terms. A pilot that does not beat the baseline is a valid result: you learn which assumption failed and avoid rolling out a change that would have cost money.

Ownership and handover

The product owner owns the measure and the rollout decision; the engineering owner owns the code and the evaluation harness. Merchandisers keep control of business rules, which sit in configuration rather than inside a model.

When to choose something else

If the problem is an existing knowledge assistant giving wrong answers, use RAG quality rescue. For a retrieval system beyond product search, see RAG and enterprise search engineering. If your support team needs to adopt AI in its daily workflow rather than a built system, see AI enablement for customer support. If you sell software to retailers, AI product delivery for B2B SaaS is closer. Other sectors are on the industries overview.

Questions buyers ask

Should we use generative AI or conventional machine learning?

It depends on the problem. Ranking products for a query or a shopper is a well-understood retrieval and ranking task, and a conventional model is usually cheaper, faster and easier to evaluate. Generative models earn their place in attribute extraction from messy supplier text, query understanding for long natural-language searches, and support answers drawn from policy content. The opportunity brief makes this call per problem.

How will we know whether the pilot helped?

We agree the baseline before building. Offline, that is relevance on a labelled query set your merchandisers trust. Online, it is a controlled test on a metric you already report, such as search exits, add-to-basket from search or contact rate. If you do not have the traffic for a meaningful online test, I say so, and the decision rests on offline evidence plus a staged rollout.

Our event tracking is unreliable. Is that a blocker?

Not for starting, but it shapes what we can claim. Recommendations depend on clean events and de-duplicated identities more than on the model. If tracking is broken, the brief records it as the first dependency, and the pilot may start with search or catalogue work that relies less on behavioural data.

Can you work with our search vendor rather than replacing it?

Usually, yes. Many improvements, such as synonyms, attribute filters, query rewriting and re-ranking, sit in front of or alongside an existing engine. Replacing a vendor is a separate commercial decision; the pilot can give you evidence for it but does not assume it.

What do you leave behind?

The pilot code in your repositories, the labelled evaluation set and harness, the opportunity brief, the test results and a rollout plan with fallback behaviour. Your engineers pair with me during the build so they can extend it, rerun the evaluation after each catalogue or model change, and decide the next item on the opportunity brief without me.

Related engagements

Industries · Regulated operations

Regulated operations

We need useful AI workflows in a controlled environment. How do we translate agreed policies into technical and review controls?

You get:Policy-to-workflow control mapping, evidence capture and pilot approval dependencies

Industries · B2B SaaS

B2B SaaS

We need an AI feature that works across customer accounts without breaking tenant isolation or unit economics. What belongs in the delivery plan?

You get:Tenant-aware feature architecture, task evaluation and per-customer cost model

AI delivery · Retrieval

RAG and enterprise search

We need an internal assistant over private documents. Who can build retrieval that respects access rights and shows its sources?

You get:Retrieval and knowledge-system implementation

Department enablement · Customer support

Customer support

Can our support team use knowledge assistants while preserving access controls, current answers and escalation to people?

You get:Permissioned support-assistance pilot, answer-quality sample and escalation playbook

Industries · Professional services

Professional-services firms

How should our firm sequence AI adoption across client confidentiality, partner review and billable delivery work?

You get:Firm-level use-case portfolio, client boundary model and governed pilot sequence

Industries · Education operations

Education operations

We need help with administrative research and communications, not automated student selection or grading. What workflow fits?

You get:Staff workflow pilot, source checks, restricted-data handling and human approval

Further reading

Describe what needs to work

A short, non-confidential description is enough to start. I read every brief personally and reply within two business days, including when the answer is that I am not the right fit.

Step 1 of 2 · The basics