Wayfair (Extern)

Deployed real-time agentic decisioning pipelines and LLM inference infrastructure on AWS for enterprise catalog workflows.

Overview

Engaged as AI Engineer & Technical Analytics Consultant through Extern's Wayfair externship. Focused on translating ambiguous enterprise catalog problems into end-to-end ML solutions — scoping deliverables, shipping agentic decisioning pipelines, and operating them with MLOps discipline.

The Challenge

Wayfair's SKU onboarding workflows involve messy, ambiguous catalog problems that resist clean rule-based automation. The challenge was scoping these fuzzy enterprise requirements into concrete ML deliverables, then deploying inference infrastructure that could operate in both batch and real-time modes while keeping inference cost and latency under control.

The Solution

I scoped deliverables with client stakeholders and deployed real-time agentic decisioning pipelines achieving 95% task completion accuracy across SKU onboarding workflows. I designed batch and real-time LLM inference pipelines on AWS (S3, Lambda, SageMaker) that integrate foundation models with structured APIs and vector databases, and built MLOps monitoring for model quality and drift — reducing redundant processing by 35%. I balanced retrieval-augmented techniques against custom LLM approaches to optimize the precision/recall/cost tradeoff, and communicated architectural decisions and model performance findings to client stakeholders in weekly reviews.

My Thoughts

The most valuable lesson from this engagement has been translation work — taking a business stakeholder's vague sense of what's broken and turning it into a scoped, measurable ML deliverable. Production AI consulting rewards clarity of communication as much as technical depth; the weekly stakeholder reviews have been as instructive as the engineering itself.

Key Achievements

• Deployed real-time agentic decisioning pipelines achieving 95% task completion accuracy
• Designed batch and real-time LLM inference pipelines on AWS (S3, Lambda, SageMaker)
• Reduced redundant processing by 35% through MLOps monitoring for model quality and drift
• Integrated foundation models with structured APIs and vector databases for enterprise workflows
• Communicated architectural decisions and model performance tradeoffs to client leadership weekly