Frequently Asked Questions

I'm building across the GenAI stack: RAG pipelines that let users chat with their documents, agentic workflow engines that autonomously decompose and execute complex tasks, and AI-augmented development tools using MCP and Claude. I'm especially interested in the applied side — taking LLMs from prototype to production with proper error handling, evaluation, and scalability.

It was a natural evolution. My 5+ years building production systems at Expedia, Tekion, and BYJU'S gave me the engineering foundation that most AI projects desperately need — reliability, scalability, and the ability to ship. When LLMs became powerful enough for real products, I had the infrastructure skills to build around them. Now I combine AI/GenAI capabilities with production engineering to build things that actually work at scale.

For LLM applications, I typically use Python with LangChain for orchestration, OpenAI or Claude APIs for inference, and vector databases (Pinecone, ChromaDB) for retrieval. The frontend is usually React or SvelteKit with TypeScript. For agentic systems, I work with tool-use patterns, MCP (Model Context Protocol), and custom agent loops. Infrastructure-wise, I focus on proper evaluation pipelines, streaming responses, and graceful fallback handling.

Mornings are for deep AI work — building features, experimenting with prompts, and evaluating model outputs. At Expedia, I work on the Workflow Orchestration Platform which is the infrastructure backbone for automation. Afternoons shift to collaboration: code reviews, architecture discussions, and writing. Evenings often involve working on personal AI projects or writing blog posts about GenAI topics.

Building a RAG system that could handle ambiguous queries across heterogeneous document types. The challenge wasn't just retrieval accuracy — it was making the system reliable enough for production use. I implemented hybrid search (semantic + keyword), re-ranking with cross-encoders, and a citation system that traces every answer back to source documents. The result was a system users could actually trust.

I'm selectively open to AI collaborations — especially building GenAI products, LLM applications, or helping teams adopt AI effectively. My role at Expedia Group is my primary focus, but I actively collaborate on open-source AI projects and enjoy technical discussions about AI architecture. The best way to reach me is through my contact page at umesh-malik.com/contact.

Start with the fundamentals: Andrej Karpathy's YouTube lectures for understanding how LLMs work, the LangChain and LlamaIndex docs for building applications, and Anthropic's prompt engineering guide for practical prompt design. Read research papers on arxiv — start with the original Transformer paper (Attention Is All You Need) and the RAG paper. Most importantly, build things. Ship a RAG app, build an agent, deploy something real. I cover practical AI topics on my blog.

AI is fundamentally changing how we write code, but it's not replacing engineers — it's amplifying them. I use Claude and Cursor daily and my velocity has increased dramatically. The engineers who will thrive are the ones who understand both AI capabilities and production engineering. Knowing how to prompt an LLM is useful, but knowing how to build reliable systems around AI outputs is what creates real value. That's exactly the intersection I focus on.

5+ years of production engineering at three enterprise companies. At BYJU'S, I built payment systems processing $10M+ monthly with 99.9% uptime — promoted from Associate to Module Lead in 8 months. At Tekion Corp, I rebuilt the F&I module serving thousands of dealerships. At Expedia Group, I lead the Workflow Orchestration Platform. This foundation in scalable, reliable systems is what I now apply to AI products.

Three reasons: performance, developer experience, and AI-readiness. SvelteKit compiles to vanilla JavaScript with no runtime overhead — perfect for a content-heavy site. I also built it to be AI-friendly: structured data, llms.txt for LLM crawlers, RSS feeds, and an AI summary page. It's a portfolio that works for both human visitors and AI systems indexing the web.