Shiva Deepak

I prefer agentic systems with explicit state over one-shot prompt chains. The useful part is not that a model can answer; it is that a system can decide whether to retrieve, clarify, use a tool, refuse, or continue a workflow.

The most underrated path in retrieval systems is the one that does not answer. Good AI infrastructure needs refusal paths, ambiguity handling, memory boundaries, and traces that make failure cases inspectable.

I run ablations on my assumptions. In Agentic-RAG, heavier reranking added latency and complexity without improving the benchmark. That kind of result matters because evaluation is how systems stay honest.

I optimize for systems I can reason about, debug, and evolve long-term.