AI Engineer: LLM Platforms and Agentic Systems

Jagadeesh V

AI Engineer LLM Platforms & Agentic Systems

Hartford, CT +1-315-***-**** *************@*****.***

PROFESSIONAL SUMMARY

AI Engineer with around 5 years of engineering experience, including 3 years shipping production LLM platforms at Mayo Clinic and Stripe. Architected multi-agent orchestration systems (LangGraph, AWS Bedrock) processing 8K+ daily requests at 99.8% uptime and P95 latency under 2.5s; built eval-driven RAG pipelines improving Recall@5 by 17pp; and drove 32% inference cost reduction through intelligent model routing. Specializes in agentic systems, RAG infrastructure, and LLMOps in regulated healthcare environments.

TECHNICAL SKILLS

Languages: Python (4 yrs professional, primary — async/await, FastAPI), SQL, Java, JavaScript, Bash

Agentic AI: LangGraph, LangChain, MCP (Model Context Protocol), tool-calling, human-in-the-loop, multi-agent orchestration, ReAct planning

LLMs & APIs: OpenAI APIs (GPT-4, GPT-4o), Anthropic Claude (3.5/3.7), AWS Bedrock (Titan, LLaMA 3), Google Gemini; multimodal inputs via GPT-4o Vision and Claude Vision (documents, images, charts)

RAG & Vector Search: Pinecone, FAISS, pgvector, hybrid retrieval (BM25 + dense), cross-encoder reranking, HyDE, embedding drift detection

Evaluation & Obs.: RAGAS, LangSmith, Langfuse, LLM-as-judge, RAG triad metrics, OpenTelemetry tracing, Recall@K, NDCG

MLOps & Serving: AWS SageMaker, vLLM, Docker, Kubernetes (EKS), Terraform, MLflow, FastAPI, GitHub Actions, Jenkins

Cloud — AWS: Bedrock, Lambda, S3, API Gateway, EKS, Step Functions, CloudWatch, KMS, EC2, IAM

Data Engineering: Apache Spark/PySpark, Airflow, Snowflake, Pandas, PostgreSQL, MySQL

Structured Outputs: Pydantic v2, JSON-schema enforcement, function-calling reliability, retry/fallback logic

Security & Govern.: HIPAA-aligned deployments, RBAC via IAM + API Gateway authorizers, OWASP LLM Top 10, NIST AI RMF, EU AI Act high-risk obligations (effective Aug 2026)

PROFESSIONAL EXPERIENCE

Mayo Clinic · Rochester, MN Jan 2025 – Present

AI Engineer – LLM Platforms & Agentic Systems

•Platform Scale: Engineered and owned the enterprise LLM platform serving 6 clinical and operational teams (~300 users), processing 8K+ daily requests at 99.8% uptime and P95 agent-response latency of 2,400ms across 400K+ indexed documents — including multimodal inputs (clinical charts, lab images, and scanned reports via GPT-4o Vision) — establishing the reliability baseline all downstream AI workloads are built on.

•Architecture Decision: Reduced new-tool integration effort by ~25% by evaluating LangGraph, CrewAI, and the Claude Agent SDK across state management, checkpointing, and observability criteria — selecting LangGraph as the primary orchestration layer for deterministic execution and standardizing tool and data access via MCP, while reserving lightweight prompt chains for sub-2-turn automation tasks.

•Eval-First Release Gate: Caught 35% of quality regressions before production and reduced post-deployment incidents by 28% by building a pre-release eval gate using RAGAS, LangSmith, and a custom LLM-as-judge harness — scoring faithfulness, answer relevance, and context relevance against a 450-example golden test set before every model or pipeline update.

•Inference Cost Optimization: Cut monthly inference spend by 32% (~$6K/month) by routing complex multi-step reasoning to GPT-4 while serving high-volume classification and extraction workloads on self-hosted LLaMA 3 via vLLM — combining prompt caching and context compression to increase throughput without measurable quality regression on RAGAS faithfulness scores.

•Security & Compliance: Passed 3 HIPAA regulatory audits with zero violations across 12 production LLM deployments by engineering a secure inference layer (API Gateway, Lambda, KMS, CloudWatch) with automated PII redaction, prompt-injection defense, and IAM-enforced RBAC — with compliance posture mapped against NIST AI RMF and EU AI Act high-risk system obligations (effective Aug 2026) for cross-jurisdictional enterprise readiness.

•Production Observability: Reduced mean time to detect model-degradation events from ~3 days to under 2 hours — catching 65% of issues before user impact — by instrumenting the full agent execution graph with OpenTelemetry spans and Langfuse traces, surfacing per-node latency, tool-call failure rates, and embedding-drift signals in real-time dashboards.

•Infrastructure & CI/CD: Enabled 3 more frequent model releases by reducing deployment cycle time from ~5 days to under 2 hours — automating infrastructure provisioning and zero-downtime rollouts via Terraform, GitHub Actions, and Jenkins with environment-parity guardrails.

•Technical Leadership: Raised team engineering bar through 20+ GenAI design reviews and hands-on mentorship of 4 mid-level engineers on agentic system design, eval methodology, and LLMOps best practices — partnering with clinical stakeholders to translate AI capabilities into measurable operational outcomes.

Stripe · South San Francisco, CA Jul 2023 – Dec 2024

AI Engineer

•RAG Architecture: Improved Recall@5 from 0.67 to 0.84 (17pp gain) across a 600K-document corpus by standardizing on Pinecone over FAISS — chosen for managed scalability and metadata filtering — then introducing hybrid BM25 + dense retrieval with cross-encoder reranking, which reduced user-reported wrong-answer support tickets by 35%.

•Fine-Tuning: Reduced RAGAS hallucination scores by 28% and improved response quality metrics by 22% by fine-tuning foundation models (Hugging Face Transformers, Amazon Titan via Bedrock) on 18K domain-specific examples using LoRA/QLoRA — validated against held-out evaluation benchmarks before every production promotion.

•Evaluation Harness: Reduced pre-release review time by 40% by building a reusable RAGAS evaluation harness against a 350-example golden test set — establishing quality regression checks as a mandatory release gate for all RAG pipeline and model changes.

•ML Deployment: Shipped ML inference pipelines on AWS SageMaker and Lambda serving 15K requests/day at 99.7% uptime, exposing model capabilities as FastAPI microservices with Pydantic-validated structured outputs and retry logic consumed by 3 downstream teams.

•NLP & Classification: Saved ~22 analyst-hours/week and reduced manual ticket triage by 45% by shipping text-classification and sentiment models using Hugging Face Transformers to automate support-ticket routing across 8 product categories.

•MLOps: Compressed model-iteration cycle from ~7 days to 2 days by containerizing deployments with Docker and Kubernetes and enforcing experiment tracking in MLflow — enabling faster A/B testing and reproducible releases across the ML team.

HCL Technologies · Hyderabad, India May 2021 – Jul 2022

Data & Backend Engineer

•Reduced downstream integration time by 35% across 5 internal consumers by building RESTful APIs in Python and Java with consistent schema contracts and versioning.

•Cut data-preparation time by 50% by developing structured PySpark and SQL ETL pipelines transforming raw transactional data into clean, analysis-ready datasets.

•Improved downstream model input quality by applying spaCy-based NLP preprocessing to extract and normalize domain-specific features, reducing manual data cleaning by 40%.

•Delivered the team's first production ML use case by building and validating a scikit-learn multi-class classifier achieving 82% baseline accuracy, enabling handoff to senior engineers for production deployment.

EDUCATION

M.S., Computer Science · University of Central Missouri, USA July 2022 - July 2023

B.S., Computer Science · JNTUK, India June 2018 - Apr 2021

CERTIFICATIONS

•AWS Certified AI Practitioner

•AWS Certified Developer – Associate

•AWS Certified Data Engineer – Associate

•Responsible AI Specialization — DeepLearning.AI

•AI Governance: Policy, Regulation & Corporate Strategy — MIT xPRO

•AI Risk Management Framework (AI RMF) — NIST

Contact this candidate