Generative AI for the Wall Street Journal

The brief

NewsCorp wanted to move generative AI from research demos into production capabilities for the Wall Street Journal — document understanding, retrieval-augmented generation, and session-aware AI interactions that could meet real production SLAs.

What I built

LLM evaluation and selection

• Led the design and development of Python/Go APIs tailored to Generative AI business use cases including document understanding and retrieval-augmented generation.
• Evaluated and benchmarked multiple LLMs — Raven, Falcon7b, LLaMA, and CodeLLaMA — assessing latency, cost, output stability, and hallucination risk for real-world deployment.
• Conducted innovative experiments with fine-tuning open-source LLMs and diffusion models (Stable Diffusion), exploring the memory-efficient LoRA technique.

Retrieval systems for production

• Designed vector-based retrieval systems using Pinecone, FAISS, and ChromaDB, implementing cosine similarity and retrieval strategies aligned with production SLAs.
• Pioneered the use of cosine similarity search techniques for efficient data retrieval and state-of-the-art RAG.

Session-aware AI services

• Built FastAPI services with DynamoDB-backed session state, supporting long-running AI interactions.
• Deployed microservices using ECS, GitHub Actions, and Lambda functions to take processing off the core API server.
• Wrote unit and integration tests to validate AI pipelines and prevent silent regressions.

Accessibility experimentation

• Led the first experimentation for screen reading of Wall Street Journal articles, aimed at helping users with disabilities engage with WSJ content.