RLGPT — Enterprise AI Assistant at Scale

Overview

Ralph Lauren's internal knowledge landscape was fragmented — employees across 9,000+ global roles navigated a maze of SharePoint sites, email threads, and support tickets just to find answers to routine questions. RLGPT changed that. Built from zero to production in 4 months, it became the single intelligent interface employees could query in plain English to find answers, locate assets, understand processes, and get things done — without waiting on a human to respond.

The system wasn't a chatbot bolted onto existing infrastructure. It was a full RAGRAGRetrieval-augmented generation. The model answers from documents fetched at query time, not only from its training. (Retrieval-Augmented Generation) platform, purpose-built for enterprise knowledge: grounded in internal documentation, aware of organisational context, and fast enough to respond in under a second even under concurrent load.

Problem Statement

Before RLGPT, knowledge lived in silos. New employees spent weeks navigating internal wikis. Support teams were flooded with repetitive queries. Design and merchandising teams wasted hours manually searching image and asset libraries for relevant references. There was no single source of truth — just people asking other people, and institutional knowledge that left when employees did.

The Brief

Build an AI system that could ingest, index, and make queryable the full breadth of Ralph Lauren's internal knowledge — documents, policies, asset libraries, process guides — and surface the right answer to the right person in seconds.

My Role

End-to-end ownership. I designed the system architecture, built the Python/FastAPI backend, implemented the RAG pipeline, developed the React admin interfaces, handled Kubernetes deployment on AKS, and wrote the Terraform IaC modules that reduced ML environment provisioning time by ~72%. I also coordinated with HR, IT, legal, and design stakeholders to ensure the platform met compliance and access control requirements across regions.

Key Features

• Grounded AI Answers: Natural language query interface with contextual, grounded responses — not hallucinated answers
• RAG Architecture: RAG pipeline over internal documentation: ingestion, chunking, embedding, vector retrieval, and generation
• Semantic Search: Embedding-based semantic search across text and image asset libraries returning results in under a second
• Admin Control: React admin dashboard for content management, model configuration, usage analytics, and access control
• RBAC: Role-based access control — employees see only content relevant to their function and region
• Global Scaling: Multi-tenant architecture supporting global teams across NA, EMEA, and APAC
• Robust Orchestration: Horizontal autoscaling on AKSAKSAzure Kubernetes Service, Microsoft’s managed Kubernetes. to handle concurrent load across time zones
• Infrastructure Automation: Reusable Terraform IaCIaCInfrastructure as code. Servers, networks and permissions defined in versioned files instead of clicked together. modules enabling new ML environment provisioning in ~72% less time

Technical Architecture

• Ingestion layer: document parsers handling PDF, DOCX, HTML, and image metadata
• Embedding layer: text embeddings via Azure-hosted models, stored in a vector database
• Retrieval layer: cosine similarity search returning top-k relevant chunks per query
• Generation layer: LLMLLMLarge language model, the kind of model behind ChatGPT and Claude. call with retrieved context injected into prompt, with guardrails to prevent hallucination
• API layer: Python/FastAPI serving all frontend and integration requests
• Infra: Azure Kubernetes Service (AKS), Docker, Terraform, CI/CD via GitHub Actions
• Frontend: React with component-based admin UI, real-time usage dashboards

Challenges and Learnings

The hardest problem wasn't the ML — it was trust. Enterprise users are unforgiving of wrong answers. The system needed to know when it didn't know, surface its sources, and never confidently produce incorrect information. This led to significant work on retrieval quality, prompt engineering, and confidence thresholding before the answer was surfaced to the user.

Scaling to 9,000+ users across time zones with sub-second response times required careful Kubernetes configuration — horizontal pod autoscaling tied to request queue depth, not just CPU. Kubernetes was the right call over serverless here because the model inference components needed warm instances, not cold starts.

Coordinating content ingestion across business units was also a human problem as much as a technical one. Getting stakeholders to trust the system with their knowledge, and to maintain it, required as much relationship management as engineering.

Outcome

RLGPT became Ralph Lauren's primary internal AI interface globally. Manual support ticket volume dropped significantly for knowledge-related queries. Design teams reported finding assets in seconds that previously took 20–30 minutes to locate. The platform became the reference point for future AI investment discussions at the organisational level.

Won the Ralph Lauren x Dataiku AI Hackathon (2024). Received the Extra Miler Award (February 2025) in recognition of the platform's organisational impact.