Amogh User

I build and ship production AI features end-to-end, with a focus on LLM application engineering—prompt design, metadata grounding, evaluation, and observability.

At Deloitte USI, I architected a supervisor-agent system for a leading US bank serving 4,000+ users, using privilege-scoped sub-agents and tool calls with per-user object permissions. I also engineered prompt templates that return structured JSON for UI-ready typed actions supporting banker daily summaries and task creation.

I strengthened reliability through RAG-style grounding over CRM data and integrated APIs to retrieve existing client PDFs for domain-specific outputs. I improved latency from 14,000ms to ~300ms (97.8%) by re-architecting bulk synchronous PII masking into async background processing.

To keep agent behavior safe and production-ready, I implemented guardrails, output schema validation, and human-in-the-loop review before writes in a regulated banking environment—along with model feedback, tracing, and dashboards for continuous improvement.