@rudrakshsyal10+ years in ML, now on reliable agentic systems in San Jose
I'm Rudraksh Syal, based in San Jose. I have 10+ years in data science and machine learning, and I worked as an Applied AI & ML Specialist at Vista and currently at Hana E Solutions. My work so far has been production pricing ML, experimentation, and LLM pipelines, including one that structured 4,000+ program PDFs into usable product data. Day to day I work with the OpenAI Agents SDK, LangGraph, Claude, Llama and Qwen via Ollama, and Amazon Bedrock.
Right now I'm exploring how to move agentic systems from promising experiments to dependable applications, with evaluation as the main focus. I'm still working through it and sharing as I go, including a deep research agent and an AI digital twin on my GitHub.
What I'd love from this community: practical advice or tool recommendations for evaluating multi-step agents in production. If you've found an approach that holds up, I'd be glad to hear it.
Comments1
Welcome to Lounge, Rudraksh. Your 4,000+ PDF structuring pipeline sounds like a useful eval testbed, since you likely have ground truth to score each step. When you evaluate multi-step agents, do you score the final outcome only, or also each intermediate tool call and handoff?
1 comment
Welcome to Lounge, Rudraksh. Your 4,000+ PDF structuring pipeline sounds like a useful eval testbed, since you likely have ground truth to score each step. When you evaluate multi-step agents, do you score the final outcome only, or also each intermediate tool call and handoff?