searchHub
German search optimization SaaS company. Helps e-commerce companies improve onsite search relevance and run statistically valid A/B experiments. Also open-sourced the Open Commerce Search Stack (OCSS) for Elasticsearch-based retail search.
Notable Contributions
A/B Testing Methodology — identified two major pitfalls in onsite search experimentation:
- Session-level randomization with purchase KPIs causes carry-over effect underestimation
- Session-based basket attribution mixes search-driven and non-search purchases — measuring wrong things
OCSS — Elasticsearch multi-analyzer pattern for retail:
- Multiple sub-fields: minimal, standard, shingles, ngram, numeric patterns
- Boolean similarity instead of TF-IDF for product data
- Query relaxation pipeline (3 queries: exact → lenient → ngram)
SmartQuery — external spell-checker and query caching layer for Elasticsearch
Scalable relevance evaluation — an architecture for running LLM judgments at a scale where doing so naively is economically impossible. searchHub reports upwards of 10 billion searches annually across its customer base, producing hundreds of billions of query-document pairs. The approach combines behavioral pruning, Staged Judging, and Knowledge Distillation into small CPU-servable models. See Towards Scalable Relevance Engineering.