Recommenders at Seznam scale
Discovery and ranking across 50M+ requests/day: clickbait mitigation, long-tail, an experimentation framework.
For e-commerce
AI that moves store metrics, not slideware
For heads of e-commerce, product leads, and founders of mid-size shops who have tried AI and want it in production.
Situations
When shops call me in
- Support tickets are 60% ‘where’s my order’.
- Product descriptions don’t scale past a few hundred SKUs.
- We bolted on a chatbot and don’t trust what it tells customers.
- Traffic from ChatGPT and Perplexity is rising and we’re invisible there.
- We have recommendation and search data and aren’t using it.
Use cases
By capability
Where each capability actually lands in a store: the result to expect, and what it needs from your catalog and data.
RAG & assistants
-
A support assistant that answers about any product, order, or policy from your catalog and knowledge base.
-
A merchandising assistant over the catalog for the team: "which SKUs are missing size data", "what do we say about X".
Chatbots
-
Storefront support and guided product discovery, with a clean hand-off to a human and multilingual where you sell.
Agents & workflow automation
-
Generate product descriptions from PIM data and draft review responses at catalog scale. Triage returns and keep feeds and PPC tidy in the background.
Classic ML
-
A recommender or a scoring model where an LLM would be too slow or too expensive: product recommendations, churn and propensity scoring, demand forecasting.
AI observability & evals
-
"Is our chatbot actually helping, or hallucinating prices?": eval sets, online monitoring, guardrails, A/B.
Timely
What's changing
Heads up
EU AI Act transparency rules, since August 2026
If your site uses a chatbot or shows AI-generated content, limited-risk transparency obligations have applied since 2 August 2026. Tell people they are talking to a bot, and look closely at synthetic images or video that could pass for a photograph. For most shops that is the whole of it. You are not a high-risk system.
An audit engagement includes a readiness pass: what applies to you, what to disclose, where the gaps are.
This is help getting ready, not legal advice. For a binding assessment, talk to a lawyer.
Note
Being found when buyers ask an answer engine
When someone asks ChatGPT or Perplexity for a product recommendation, your store is either in the answer or it isn’t, and that traffic is growing while it stays uncredited in most analytics.
Answer-engine optimisation is mostly retrievability: structured catalog data, schema.org markup, content that states things plainly, and pages a model can actually read. I audit where you stand and fix what is in your control.
Expectations
How this actually goes
I work with Shoptet, Shopify, and custom stacks. Most AI failures in e-commerce are data and architecture failures, not model failures, so the first week is an honest look at your catalog, orders, and events.
Track record
Systems I’ve built
Prior in-house work close to store problems: recommendation and ranking at media scale, and an email platform built on segmentation, deliverability, and A/B from the core.
Email Machine, concept to production
Co-founded and scaled an email platform from one retailer's internal tool to a SaaS running 100M+ emails/month for 196 brands.
Packages
Three ways to start
Fixed scope, fixed fee: quoted after the intro call.
AI Discovery Sprint
For Teams unsure where AI actually pays off.
Scope
- Your situation, your data, your constraints: reviewed
- 2–3 candidate use cases, ranked by impact and effort
- Approach decision per case (LLM API vs a model, RAG vs a recommender, build vs buy)
- Success metrics defined up front
- A one-page architecture sketch
You get
- A prioritised use-case shortlist
- An approach recommendation with the reasoning
- Success metrics and an effort/impact estimate
- The architecture sketch
Outcome You know which AI work to do first, why, and what 'good' looks like.
Fixed fee, quoted after the intro call.
Book the intro callPrototype Build
For A chosen use case that needs to exist and be measured.
Scope
- A working prototype, integrated with your systems
- A baseline evaluation: offline metrics, an A/B or sanity checks
- Documentation and a next-steps roadmap
- A handover to your team
You get
- The integrated prototype
- The evaluation and its numbers
- Docs + roadmap
- A working session to hand it over
Outcome One AI feature in front of real users, with numbers you trust and a team that can run it.
Fixed fee, quoted after the intro call.
Book the intro callAudit & Fix
For An existing RecSys / ML / LLM feature that underperforms or isn't trusted.
Scope
- An audit of the metrics, the data, and the architecture
- A quick-wins list
- An eval harness so regressions get caught
- An improvement roadmap with estimated impact
- For e-shops: an EU AI Act readiness pass (help, not legal advice)
You get
- The audit findings
- The quick-wins list
- The eval harness
- The roadmap
Outcome You know why it underperforms, what to fix first, and how you'll know it improved.
Fixed fee, quoted after the intro call.
Book the intro callFAQ
Common questions
Do you work with Shoptet and Shopify?
Yes, both, plus custom stacks. Shopify has the richer API surface, so assistant and automation work moves faster there; Shoptet is workable through its API and exports. Either way the first question is what your catalog and order data actually look like, not the platform badge.
Will an assistant have access to live order data?
Only through a read path you control, and only the fields it needs. A “where’s my order” answer needs order status and tracking, not a customer’s full history. Access runs through your API or a scoped view, is logged, and is agreed in writing before any build starts.
How do you handle the AI Act chatbot disclosure?
It’s part of the readiness pass in an audit. Since 2 August 2026, limited-risk transparency rules mean you tell people they are talking to a bot, and you look closely at synthetic images or video that could pass for real. I map what applies to your setup, draft the disclosure wording, and flag the gaps. This is help getting ready, not legal advice: a binding assessment is a lawyer’s job.
Can you improve how we show up in ChatGPT and Perplexity results?
Partly, and honestly: no one controls what an answer engine says. What moves the needle is making your catalog and content easy for a model to retrieve and cite: clean product data, structured schema.org markup, plain copy, and pages that are not locked behind scripts. I audit that and fix what is in your control.
What do you need from our catalog to start?
Read access to your products, orders, and events: a sample export is enough for the first week. That week is an honest look at data quality: missing attributes, inconsistent categories, event gaps. Most e-commerce AI failures trace back to this, so it comes before any model work.
Tell me where your store is stuck.
Thirty minutes, no pitch. Your data, your stack, whether this is the right problem.