Subramanya

R&D Software Engineering Specialist (AI Backend Engineer)

Collapsed multi-step PLM engineering workflows into single natural-language requests for 500+ engineering users across aerospace and automotive programs, by architecting an autonomous natural-language interface on LangGraph (ReAct) with role-specific agent competencies that mutate engineering objects and transition lifecycle states through governed production APIs.. Eliminated the over-privileged-agent failure class, measured by zero unauthorized mutations in production, by building Model Context Protocol (MCP) tool integrations with cross-application RBAC — authenticated user roles propagate into agent context, and a deterministic execution layer validates every permission server-side before any downstream API mutation fires.. Kept regulated aerospace and automotive data inside the enterprise boundary, measured by zero third-party LLM egress, by deploying self-hosted enterprise LLMs (MLI) across an air-gapped private enterprise cloud — Kubernetes-orchestrated inference services fronted by vLLM / Triton serving engines using continuous batching to sustain GPU utilization across 500+ concurrent users.. Blocked agent quality regressions from ever reaching production users, measured by automated pass/fail gates on every release candidate, by building an LLMOps evaluation pipeline — golden datasets of validated PLM interactions, automated scoring against a committed baseline, and hallucination regression tests that fail the build when agent accuracy degrades.. Delivered sub-200ms time-to-first-token and resumable multi-turn agent sessions, by streaming reasoning traces and tokens over Server-Sent Events with conversation state and session persistence in PostgreSQL surviving pod restarts and rescheduling.. Delivered a 60% retrieval speedup in production on a globally deployed configuration-and-release framework serving Porsche, BMW and Safran, measured by JMeter against an identical pre/post load profile, by implementing multi-level Redis caching and refactoring JPO/MQL query patterns — as sole backend owner of 40+ RESTful endpoints modelling hierarchical product-lifecycle data.. Sustained 99.9% uptime SLAs across Tier-1 client deployments, measured by 180+ distributed-system defects diagnosed and resolved before production impact, by leading JMeter performance and capacity audits across high-load production scenarios.. Achieved zero critical security regressions across 12 months of senior-level code review, by hardening API surfaces with OAuth2/JWT token flows, method-level authorization, and structured audit logging.