Leon Raymer

I’m a Senior ML Engineer with 4+ years of production experience delivering end-to-end NLP and LLM solutions at scale. I’ve built and owned RAG systems, LLM fine-tuning pipelines, and evaluation infrastructure serving millions of users. I translate business requirements into measurable outcomes, from support automation to content intelligence and conversion optimization.

At Yandex, I implemented a custom CUDA-accelerated loss function within the PyTorch training kernel, reducing GPU memory overhead by 9% and cutting cloud compute costs by ~$14K/month for domain-specific LLM fine-tuning. I also built an offline LLM evaluation framework with an automated benchmark suite of 3K labeled examples, LLM-as-a-Judge scoring, and Slack alerting on metric degradation—reducing regression detection time from 2 weeks to 4 hours. In production, I shipped a review summarization feature, aggregating 50–500 reviews per product with a map-reduce approach and achieving a 3.41% relative increase in add-to-cart rate on 2M users via A/B testing.

Earlier at Yandex, I designed and deployed a RAG-based customer support automation system using BM25 retrieval (Elasticsearch), a cross-encoder reranker, and a fine-tuned, grounded LLM—automating 34% of incoming support tickets with CSAT of 4.1/5.0 and keeping hallucination rate at 2%. I replaced a legacy BERT classifier with an instruction-tuned LLM (SFT + LoRA) for ticket routing across 20+ categories, improving accuracy from 81% to 93%, especially on long, multi-issue tickets. I’ve mentored interns and junior engineers through onboarding, code reviews, and production deployments, and I’ve helped lead design reviews to align engineering effort with product roadmaps and business priorities.