Suvom Karak

Architected a shared agent runtime on LangGraph and FastAPI used by 6 product teams, enabling supervisor/worker agents with typed state, PostgreSQL checkpointing, and durable execution. Designed an agent governance layer that routes tool calls through policy-as-code with default-deny access, per-turn budgets, circuit breakers, and idempotent execution to prevent duplicate writes. Exposed agent capabilities as MCP servers with OAuth2.1 and JWT-scoped RBAC, including tenant-isolated access controls and verification before execution. Built a planner–executor orchestrator (orch es trator) that decomposes requests into a dependency DAG and executes steps in risk tiers from read-only through approval-gated writes, lifting autonomous completion from 61% to 84%. Delivered production deployments on Bedrock Agent Core with sandboxed sessions and isolated tool access, reducing agent release lead time from 4 days to under 2 hours. Implemented LLM evaluation and observability with golden-set regr

Built an autonomous hiring orchestration system on LangGraph for an internal recruitment client, including a 13-node journey with human approval checkpoints covering outreach, AI screening, interviews, and end-to-end offers, reducing time-to-shortlist from 9 days to 31 hours across 3.1k candidates per month. Designed the sourcing agent team via parallel searches across LinkedIn, job boards, and an internal talent index, using a ranking agent to score candidates under a single rubric and reduce end-to-end search time from 56s to 31s. Engineered screening agents that extract evidence per competency from voice/video interviews into typed outputs and summary scoring for recruiters, moving 68% of first-round screens off the recruiter calendar. Implemented deterministic handling for offers, messages, and data lookups using idempotency keys, journey locks, and approval gates to avoid duplicates (0 duplicate offers across 14 months). Established an evaluation loop with a 220-case regression se