GenAI customer understanding and recommendation platform
Amazon · Lead Engineer, New Initiative
Overview
I was the lead engineer for a new Amazon initiative, building its next-generation customer understanding and recommendation platform from the ground up. I led the architecture and technical direction and helped grow the team, using GenAI and machine learning to understand customer behavior and improve personalization for 100M+ customers.
My work spanned long-context LLM fine-tuning, large-scale inference optimization, experimentation, causal measurement and agentic evaluation. I worked closely with scientists, product leaders and engineers across teams to shape the technical roadmap, simplify complex systems and turn research ideas into reliable, production-scale solutions.
Key contributions
- Recommendation system and business impact: Led the architecture and development of a multi-stage recommendation system combining traditional ML models with GenAI to improve personalization. Initial experiments showed estimated annualized gains of $50M+ in incremental sales and 10M+ units.
- LLM training: Productionized long-context LLM fine-tuning for customer understanding using QLoRA and distributed training. Built repeatable workflows that enabled engineers to launch training runs within the same day.
- Inference efficiency: Optimized vLLM inference across 253M customer profiles, delivering approximately $4.3M in annualized compute savings, based on measured savings per biweekly refresh.
- Agentic evaluation: Built an evidence-grounded quality system, increasing recommendation-quality scores from 62.6 to 87.4 and making evaluation decisions auditable.
- Causal measurement: Defined the engineering architecture for a shared long-term customer-value measurement platform, enabling teams to reuse customer states, embeddings and causal signals across use cases.