Konstantinos Xeras

Product manager for data, search and AI products, based in Kraków

I build working products to understand product problems.

Each project runs on a real API or platform. Each one documents the problem, the trade-offs, what I found, and how it changed version by version. I publish the reasoning along with the code.

Work

Newest first. Findings are marked confirmed or hypothesis, and sandbox, sample and synthetic data are named as such.

v0.1.0Live demo

Search Relevance Lab

BM25, sentence embeddings, open WANDS dataset

Search teams argue about relevance from a handful of example queries. This app runs one query through keyword, semantic and re-ranked search on a catalogue of 42,994 products, and scores each method against labels that people made by hand.

Confirmed on test queries

Letting any query word match lifted nDCG@10 from 0.406 to 0.704. The learned blend added 0.034 on top.

Hypothesis

Semantic search is under-scored, because only 65% of its top results have a label.

Open WANDS dataset, 192 held-out test queries, scored 16 Nov 2026. Furniture and home goods only. Free hosting can take a minute to wake up.

v0.1.1Live demo

Review Radar

Claude API, App Store and Google Play reviews

Product teams sit on thousands of app reviews spread across stores, countries and languages, and nobody reads them all. Review Radar collects them, has Claude group them into themes with a severity score, and ranks what to fix first, with the verbatim reviews behind every rank.

Product call

The AI proposes a theme list once, then tags every review against it, so counts compare across markets.

Product call

The ranking is a simple weighted score with adjustable weights. The AI does not choose the order.

Tested end to end with synthetic data only. First live-data case study in progress. Free hosting can take a minute to wake up.

v0.1.1Tested, no pilot yet

Garage Quote Agent

n8n, Claude, WhatsApp

Small car-repair garages lose customers because nobody answers "how much?" while the mechanics are under a car. This n8n automation answers price requests from WhatsApp, a web form or a webhook. AI reads the message, the garage's own price book sets the price, and the owner approves anything unclear from their own WhatsApp.

Confirmed in testing

Vague, unsafe and owner-only requests went to the owner with the reasons listed. Routine jobs with a known car were answered automatically.

Hypothesis

Most requests are routine enough to answer automatically. To be measured in a pilot.

23 of 23 test scenarios passed in n8n with sample data and stand-in services.

v1.4.1Live demo

KostasTravel

Nuitee LiteAPI, room mapping and content quality

A hotel booking site and a room-mapping explorer on the same travel API. The site shows what mapping and content quality look like to the guest who pays. The explorer shows the machinery: which supplier room names land on which master room, which rooms never get a price, and which records look alike.

Confirmed in sandbox data

533 of 533 supplier rates resolved to a master room, so coverage was fine.

Confirmed in sandbox data

The problems were precision and content. Three hotels held two master rooms under an identical name.

LiteAPI sandbox, Kraków, 40 hotels analysed in depth, 13 Nov 2026. Illustrative, not production figures. Free hosting can take a minute to wake up.

In progress

Rental Health Check

Property, Kraków

For people who rent out flats in Kraków: rental yield and profit, the flat's current value, and what selling would leave you with, side by side.

Question

Is this flat still a good investment, compared with selling it today?

Repo and demo coming soon.

What every project documents

  • ProblemWho it hurts and the question I want answered
  • DecisionsOne record per real trade-off
  • BuildA working app on a real API
  • FindingsEvidence, marked confirmed or hypothesis
  • VersionsChangelog and releases

About

I'm a product manager with 12+ years across Fortune 200/500 companies. I work on data, search and AI products, most recently on automotive parts catalogue and search.

I build these projects myself, with Claude (Anthropic) as my engineering partner. The problem framing, scope, decisions and findings are mine, and every repo says how it was built.

Focus
Search relevance, product data quality, AI products
Experience
Genuine Parts Company, Kimberly-Clark, Cisco, Stibo Systems
Education
MBA, MSc Information Systems, CPM
Based in
Kraków, Poland