← All services
Service · 09

AI Search & Recommendations

Your customers search in sentences and your catalogue answers with filters. I build natural-language search and recommendations on your catalogue, tested on real queries. Recommendations can run on the TasteRay API, which I develop as CTO.

Who it’s for: For classifieds, marketplaces, shops and services with a large catalogue: property, products, content, events, jobs.

If customers search by product name and your current search handles that well, new search will change little. I check that in the test before anything is built.

Message me on LinkedIn
time
5-day test · 3–⁠4-week implementation
format
on your catalogue and queries
outcome
before/after comparison + implementation
price
from €3,500 net · test
( 01 )

Zero results for a question that has an answer.

A user types “two bedrooms near the metro under 3,000” or “a gift for my dad who loves cooking”, and the search engine looks for words. Some queries return nothing, others return a list that misses. In 2023 I shipped mieszkAI at Morizon and szukAI at Gratka: the first natural-language property search in Poland, over 220,000 listings. I can build something similar for you.

( 02 )

What you get

  • A test set built from your real queries, including those that return zero results today
  • Current and new search compared on the same set, judged blind by your domain expert
  • Natural-language search: the query turned into filters plus semantic search over your data
  • Recommendations that explain why an item fits: on the TasteRay API or on your own engine
  • Cost per 1,000 queries and response time, calculated before anything reaches production
  • An A/B test on part of the traffic and a decision: scale, improve, or keep the current search
( 03 )

How it runs

01Week 1Test: I collect queries, build the test set and compare current search with the new one. You can stop here.
02Weeks 2–⁠3Implementation: search and recommendations on your catalogue, behind a flag and with a cost cap.
03Week 4Measurement: an A/B test on part of the traffic, the result and a decision.
( 04 )

What you need on your side

The implementation works if you have:

a catalogue export or access to its API
search query logs from recent months, without personal data
someone who knows the catalogue and can judge results
a team that will maintain the solution afterwards
( 05 )

Three examples

01

Property

“2 rooms near the metro under 3k” → location, price, size → listings

zero-result rate · listing clicks · enquiries to sellers

02

Shop

“a gift for my dad who loves cooking” → intent and budget → products ranked with reasons

search clicks · add-to-cart · returns

03

Content and entertainment

“something light for tonight, no violence” → mood and preferences → recommendations from the TasteRay API

clicks · time to choose · return visits

( 06 )

Price

from €3,500net · test

The starting price covers the five-day test on one catalogue in one language. Implementation on one search or recommendation surface is quoted after the test, as a fixed fee for a fixed scope.

  • implementation starts at €13,000 net and depends on data sources, integrations and languages
  • the test fee comes off the implementation price if you order it within 30 days
  • TasteRay bills the TasteRay API at its public prices, with no markup from me

Tell me which catalogue you want to search and how many items it has. Within 1–⁠2 business days I’ll reply with a price, or with the questions I need answered to set one.

As of: November 2026

( 07 )

Before you write.

Why the TasteRay API?

Because I know it from the inside: I develop this engine as TasteRay’s CTO. Every recommendation comes with an explanation of why it fits, and the user’s preferences are sent with the request, so you don’t need months of click history. I co-founded TasteRay, so I’ll say it plainly: if the test shows another engine or your own solution works better, that’s what I recommend.

Do we need user behaviour history?

Not to start. Natural-language search works on the descriptions in your catalogue, and TasteRay recommendations work on what the user says about their preferences. Click history helps later, when tuning the ranking.

How do we know the new search is better?

First from the test: the same set of real queries, results judged blind by your expert. Then from an A/B test on part of the traffic. A better test score doesn’t guarantee higher conversion, so the decision to scale is made after the production test.

What about model costs?

I calculate the cost per 1,000 queries before implementation and set a cap. Part of the traffic is served from cache and cheaper models; a more expensive model only goes where it improves the result.

Who maintains it afterwards?

Your team. The code lives in your repository, with documentation and the test set you can use to check every later change.

( 08 )

What’s next

You keep working search, the test set and the A/B result. When new search exposes messy catalogue data, the next step is often an Agentic Workflow Sprint, for example attribute enrichment or listing moderation.

Does your search return zero results?

I reply personally

Message me on LinkedIn