Suman Debnath is director of developer relations and product at Crusoe, specializing in AI infrastructure, managed inference, and the engineering required to run advanced models reliably. His work spans distributed storage, machine learning, visual document retrieval, open-source agents, and production AI architecture.
Debnath began his career in systems, storage, and performance engineering, with roles at Toshiba, Broadcom, and NetApp. At Amazon Web Services, he worked on Amazon Elastic File System and became a principal developer advocate focused on machine learning and generative AI. In 2024, he coauthored the launch of single-document question answering for Amazon Bedrock Knowledge Bases, enabling users to query individual documents without setting up a persistent vector database or ingestion pipeline.
He subsequently led developer relations at Anyscale, working across Ray, distributed training, fine-tuning, and inference optimization, before joining Crusoe.
- Match retrieval to the document. For scanned forms, illustrated textbooks, and image-heavy instructions, separating text, tables, and images can destroy their shared context. Debnath’s visual document retrieval workflow uses ColPali to embed complete pages as visual patches and Qdrant to identify relevant pages through late-interaction visual retrieval. He recommends this heavier architecture only when simpler text-based retrieval cannot answer the question reliably.
- Build agents from working tools. Debnath helped introduce Strands Agents, an open-source SDK that pairs capable models with callable tools while supporting Amazon Bedrock, LiteLLM, and Ollama. His implementation examples extend a functioning retrieval pipeline with spoken answers and connect Model Context Protocol servers to Manim for mathematical animations. His practical distinction: when a specific speech behavior must be dependable, configure the tool directly instead of relying on prompts.
- Evaluate realistic analytical work. He coauthored DSBC, a benchmark for data-science agents, examining ambiguous instructions, data leakage, temperature sensitivity, and tasks that combine multiple analytical operations. The research tests whether costlier agentic execution actually improves results.
- Engineer for failure. Debnath’s writing on reliable agent systems argues that tool timeouts, degraded retrieval, corrupted context, and cascading latency demand production-ready agent architecture: durable state, observability, dependency isolation, checkpointing, and recoverable execution.
