Sravan Pusuluri
Sravan Pusuluri
Generative AI Engineer
in GH ✉ ☏
Open to Remote
Ready to Relocate


who makes AI |

Building LLM inference pipelines, RAG architectures & agentic AI systems at Fortune 50 scale. Based in New York, NY.

42%
Latency Reduced
35%
GPU Memory Saved
$620K+
Infrastructure Saved
4+
Years Experience
View My Work Get In Touch
About Me

The engineer behind
the numbers.

I'm a Generative AI Engineer with 4+ years of experience building LLM inference pipelines, RAG architectures, and agentic AI systems at enterprise scale.

Currently at MetLife (Fortune 50), I work on the full stack of AI infrastructure — from quantization and GPU optimization to multi-agent orchestration and observability.

My work spans regulated industries including finance and healthcare, where responsible AI and production-grade guardrails aren't optional — they're required.

Let's Connect
📍
Location
New York, NY · Open to Remote
New York, NY — where the servers never sleep and neither do I. Open to Remote & Relocation! 🗽
🏢
Current Role
Generative AI Engineer @ MetLife
Officially: Generative AI Engineer. Unofficially: the person who stops the LLMs from hallucinating at 3am. 🤖
🎓
Education
MS Data Analytics, SUNY Albany '25
🎓 SUNY Albany, Class of 2025. Spent more time optimizing LLMs than sleeping. My thesis advisor called it 'impressive'. My mom called it 'concerning'. 😄
Technical Skills

Everything I work with,
end to end.

GenAI & LLMs
LangChain LlamaIndex Bedrock vLLM TensorRT-LLM Llama 3 Mistral RAG LoRA/QLoRA RLHF
Agentic AI
LangGraph LangSmith AutoGen ReAct ReWoo Multi-Agent AI Guardrails Responsible AI
Vector & Search
Pinecone pgvector ChromaDB Weaviate OpenSearch Semantic Search Hybrid Search
GPU & Performance
CUDA NCCL Nsight Systems INT8 Quant KV-Cache Dynamic Batching DeepSpeed MII A100/H100
MLOps & Cloud
Docker K8s SageMaker Actions Airflow FastAPI MLflow Prometheus CloudWatch
Languages & ML
Python PyTorch TensorFlow Hugging Face SQL Scala Spark XGBoost spaCy
Experience

Where I've built
real impact.

MetLife · Fortune 50
Generative AI Engineer — LLM Optimization & Inference
Sep 2024 – Present · New York, NY
Current
  • Optimized LLM inference stack using vLLM and TensorRT-LLM, cutting p95 response latency by 42% while preserving output consistency across 3B-parameter model variants in production.
  • Deployed quantized INT8 GPTQ pipelines with KV-cache streaming and head pruning, reducing GPU memory footprint by 35% on NVIDIA A100 clusters — enabling 2× model concurrency and deferring an estimated $500K in GPU hardware procurement.
  • Integrated DeepSpeed MII for multi-GPU orchestration via NCCL and CUDA Graphs, achieving near-linear scalability across 8-GPU nodes.
  • Engineered dynamic batching in Triton Inference Server, increasing GPU utilization from ~58% to 83% — avoiding an estimated $120K+ in annual cloud GPU spend.
  • Converted fine-tuned models to ONNX Runtime for edge deployments, achieving 1.7× token throughput vs. baseline PyTorch.
  • Architected RAG pipelines using LangChain and vector databases (Pinecone, ChromaDB) with measurable hallucination reduction.
  • Built production-grade inference monitoring with Prometheus + CloudWatch, enabling real-time latency alerting and automated horizontal scaling.
  • Automated CI/CD via Docker + GitHub Actions, compressing release cycles from weekly → daily and improving deployment reliability by 50%.
  • Led GPU performance investigations using Nsight Systems, reducing inference variance by 27%.
Sage Softtech
Machine Learning Engineer
Mar 2021 – Jul 2023 · India
  • Engineered production ML system on 50M+ EHR records — patient readmission risk engine improving high-risk detection by 18%.
  • Built clinical NLP pipeline using spaCy NER, boosting predictive AUC by 0.09.
  • Compressed retraining cycles from 4 hours → 45 minutes via Airflow-orchestrated pipelines.
  • Deployed production APIs (FastAPI + Docker on AWS EC2) serving 200+ physicians daily.
  • Implemented MLflow + Prometheus for full model lifecycle governance and automated data drift detection.
  • Operated within a regulated healthcare environment — the same compliance mindset now applied at MetLife.
Projects

Things I've built
that actually work.

01
Enterprise RAG Chatbot with Agentic Workflows
Multi-agent conversational AI using LangGraph (ReAct) and Amazon Bedrock, grounded in a 100K+ document corpus via ChromaDB. Implemented AI Guardrails for PII redaction and hallucination mitigation.
91% faithfulness <800ms p95 500+ concurrent
LangChain LangGraph Amazon Bedrock ChromaDB LangSmith AWS Lambda
02
Fraud Detection System — Real-Time Anomaly Scoring
Fraud detection pipeline on highly imbalanced financial transaction data using Isolation Forest and Autoencoder-based anomaly detection. Deployed as real-time scoring API on SageMaker.
94% recall Sub-second alerts
Isolation Forest Autoencoders XGBoost SageMaker CloudWatch
03
Multi-Model Time Series Demand Forecasting
Ensemble forecasting system comparing ARIMA, Prophet, and LSTM for supply chain demand. Enabled 30-day ahead forecasts with automated seasonality detection.
12% lower MAPE 30-day horizon
LSTM Prophet ARIMA Apache Airflow AWS
04
Customer Churn Prediction
End-to-end machine learning pipeline to predict customer churn using classification models. Built with full data preprocessing, feature engineering, model training and evaluation pipeline.
End-to-end ML pipeline
Python Scikit-learn Pandas XGBoost EDA
Click to view project →
Blog

Thoughts on AI & Engineering

01
How I cut LLM latency by 42% at MetLife
LLM MLOps
March 2025 · 6 min read
02
RAG vs Fine-tuning: What I learned building enterprise AI
RAG LLM
June 2025 · 8 min read
03
Why GPU memory optimization saved us $620K
GPU Infra
November 2025 · 5 min read
04
Agentic AI in production — lessons from multi-agent systems
Agents LangGraph
February 2026 · 7 min read
Featured post
How I cut LLM latency by 42% at MetLife — a production deep dive
March 2025 · 6 min read
A behind-the-scenes look at optimizing vLLM, TensorRT-LLM and INT8 quantization pipelines on NVIDIA A100 clusters at Fortune 50 scale. Covers KV-cache streaming, dynamic batching, and how we deferred $500K in GPU spend.
LLM MLOps GPU Production
42%
Latency cut
$500K
GPU savings
6 min
Read time
Read article →
Certifications

Credentials that prove
the expertise.

Amazon Web Services
AWS
In Progress
AWS Certified Machine Learning Engineer – Associate
MLA-C01
Amazon Web Services
AWS
Completed
AWS Certified AI Practitioner
AIF-C01
NVIDIA Academy
In Progress
Deep Learning Institute – Accelerating LLM Inference
NVIDIA DLI
NVIDIA Academy
Completed
AI for All: From Basics to GenAI Practice
Sep 2025
Click to view certificate
NVIDIA Certificate
Contact

Let's build something
great together.

Open to senior Generative AI / MLOps roles. Feel free to reach out via any channel below.

✉
Email
sravanpusuluri03@gmail.com
in
LinkedIn
linkedin.com/in/sravan-pusuluri
GH
GitHub
github.com/sravanpusuluri
☏
Phone
(518) 888-0259