@xzaviourrAgentic AI and evals at Microsoft Bangalore: Anshul, IIT Bombay
I'm Anshul Sanghi, a software engineer at Microsoft in Bangalore with an M.Tech in Computer Science from IIT Bombay. I build agentic AI solutions across education, health, transport and customer experience. My day-to-day mixes LangGraph, MCP tools, context pipelines and LLM evaluations with .NET APIs, React and Azure. I also use Claude, Gemini, ChatGPT, Claude Code and GitHub Copilot.
What's working: full-stack AI insight experiences that have served 7,500+ API requests, and multi-agent systems for screenplay-to-video generation and market intelligence. A Synapse-to-Kusto migration cut our API response times by 50%. What I'm still figuring out is how to turn ambiguous AI ideas into measurable, reliable products, especially through evaluation loops.
I'd love to hear from anyone who has built evals for multi-agent systems: what you measure, and which tools you trust.
Comments1
Welcome to Lounge, Anshul. For multi-agent systems like your screenplay-to-video pipeline, where outputs are subjective, do you score each agent's step separately or only the final result? I'd guess the per-step scores catch regressions earlier, but they're harder to tie to what a user cares about.
1 comment
Welcome to Lounge, Anshul. For multi-agent systems like your screenplay-to-video pipeline, where outputs are subjective, do you score each agent's step separately or only the final result? I'd guess the per-step scores catch regressions earlier, but they're harder to tie to what a user cares about.