Sandipan Bhaumik is a data and AI technical lead at Databricks and the creator of AgentBuild, a newsletter and community devoted to making enterprise AI agents reliable. He applies distributed-systems architecture to a problem enterprises routinely underestimate: keeping autonomous software observable, governed, and recoverable after deployment.
Before joining Databricks, Bhaumik worked at Amazon Web Services, helping organizations build cloud data and analytics platforms. His AWS writing on analytics-as-a-service examined multitenancy, data isolation, security, and governance across managed, customer-controlled, and hybrid architectures. Experience with financial services, healthcare, and software companies shaped his emphasis on operational accountability in regulated environments. He launched AgentBuild in 2025 to share practical approaches to production agentic systems.
- Evaluation before model selection. Bhaumik starts with measurable business outcomes, representative examples from domain specialists, and automated testing. In a retail-banking deployment, model selection followed the construction of an evaluation pipeline. Production failures feed a living evaluation dataset, while targeted test runs control evaluation costs.
- Behavioral agent evaluation. Correct answers can conceal duplicated database calls, unsafe tool choices, unnecessary retries, or expensive loops. His enterprise production framework combines deterministic checks, model-based semantic assessment, execution tracing, governed data, and clear incident ownership.
- Multi-agent systems are distributed systems. Bhaumik traced inconsistent credit decisions to a stale cache that left agents working from different customer records. His orchestration architecture distinguishes autonomous event choreography from centrally managed workflows and uses immutable versioned state, circuit breakers, and compensating transactions to prevent cascading failures and support recovery.
- Agent handoff contracts. In his analysis of failed agent handoffs, Bhaumik argues that interfaces must preserve meaning, provenance, authority, and user intent—not merely transmit correctly formatted data. A risk assessment, for example, should never become authorization to freeze an account through a downstream model’s unsupported inference.
For Bhaumik, production readiness extends to privacy safeguards, auditable execution, prompt change management, model upgrades, human escalation, and explicitly assigned operational responsibility.
