Anshuman kumar Soni

I’m a final-year B.Tech Computer Science student building LLM-powered systems that ship only after I can answer, “how do we know this works?” In my internship, I worked on multi-turn, agentic conversational pipelines using OpenAI and Claude APIs, including automated chat and audio transcription workflows.

I focus on evaluation and closed-loop debugging: testing pipeline outputs iteratively for correctness and reliability, analyzing conversation-level logs, identifying failure modes, and routing fixes back to engineers. I design prompt structures and API integration logic so agent behavior is consistent, measurable, and backed by documented scoring criteria shared with cross-functional stakeholders.

On the systems side, I build with a production mindset—FastAPI REST APIs, streaming, async Python, and observability. My projects include a memory-augmented AI backend with a hybrid retrieval scoring framework, benchmarked retrieval latency improvements, and instrumentation for performance regressions; and a self-healing backend that classifies runtime faults, triggers automated recovery, and monitors health in real time with Prometheus and Grafana.