Scaling Semantic Search with LLMs and ML

​What does it take to turn a vague recruiter query into precise results from a pool of 300M+ profiles? At Weekday, we built a high-speed semantic search pipeline combining LLMs, classic ML, and structured retrieval, all optimised for latency, cost, and retrieval depth. This talk dives into the real engineering tradeoffs behind our hybrid architecture. We’ll show how clustering, embedding normalisation, and graph-based entity linking play a central role alongside LLMs, where large models degrade into rule-based behaviour on noisy job inputs, and what it really takes to make them production-ready. We also cover how we evaluate retrieval quality using LLM-as-a-judge techniques.

About speaker:
Ayush Mittal Founder at Stealth
LinkedIn: /ayush-mittal-3b201553
Previously CTO at Lifio.ai and Led Data Science team at SharaChat.