What the work involves

This is an engineering role, not an advisory or annotation one. You own agentic systems end to end: the retrieval pipelines that feed them, the backend services and MCP/tool-calling interfaces around them, the orchestration layer (LangGraph, Google ADK, CrewAI, Claude Agent SDK or equivalent), the evaluation harness that keeps regressions visible, and the AWS infrastructure they run on. Day to day means writing and reviewing large-scale Python and SQL, shaping context windows to hold token cost down, wiring validators and deterministic fallbacks so an agent failing a policy check degrades instead of hallucinating an action, and tuning latency against agreed SLOs through caching, model routing, batching and parallel tool calls.

Evaluation is a first-class part of the job rather than an afterthought. You are expected to build offline eval sets, run continuous online evaluation, detect regressions before clients do, and instrument traces at a level that survives scrutiny in regulated enterprise environments — LangSmith, Langfuse or comparable. Adversarial test suites and validator models are explicitly named in the responsibilities, so bring examples of catching failure modes rather than only measuring happy paths.

What the screen looks for

  • Production agentic depth, not demos. At least two agentic systems you designed and took to production, with specifics on tool schemas, failure handling and what broke after launch.
  • Genuine systems engineering. 8–14 years total, strong Python at scale, plus working depth in Go, Rust, Java or C/C++ and the judgment about when it's warranted.
  • RAG from first principles. Chunking, embeddings, hybrid retrieval, reranking, response validation — and vector store choices you can defend.
  • Cloud and IaC fluency. ECS/EKS, Lambda, S3, DynamoDB, Redshift, Step Functions; Terraform or CloudFormation; mature CI/CD. Azure or GCP equivalents are accepted.
  • Client-facing clarity. You will translate business problems into technical roadmaps directly with Fortune 500 stakeholders.

Logistics

Full-time employment, based in India (Bengaluru), three days a week in office — this is not a remote or async engagement. Working hours require a daily overlap with US Eastern morning, roughly 8AM–12PM EST, which lands in the Indian evening. Compensation is not disclosed in the posting; ask for the band early in the process rather than after technical rounds. Expect a hands-on technical screen and system design discussion focused on agent architecture, evaluation and cost/latency tradeoffs.