Skip to the work

Hemanth Kiran Reddy PoluSenior AI Engineer at Mastercard

I build AI agents for companies where a wrong answer is expensive.

Since 2024, that has meant the agent and retrieval layer of Mastercard’s enterprise copilot platform. Before that, Bank of America, FedEx, Blue Cross Blue Shield and AT&T, starting with Java and SQL in 2014 and moving through machine learning to language models.

Read the workhemanthreddypolu@gmail.comLinkedIn
Portrait of Hemanth Kiran Reddy Polu
Hemanth PoluSenior AI Engineer
Now
Mastercard, since April 2024
Before
Bank of America, FedEx, BCBSA, AT&T
Focus
Agents, retrieval, evaluation, MLOps
Open to
Senior AI engineering roles

Work

Twelve years, five companies, one habit: build the version that still works after the demo. Java and data systems first, then machine learning, now agents and retrieval. Open any role for what I built and how it worked.

An enterprise copilot platform for knowledge discovery and workflow automation. I build its agent and retrieval layer: the parts that find the right information, act through approved systems, and prove an answer is grounded before anyone sees it.

What I built

  1. 1Multi-agent workflows. LangGraph and LangChain graphs with shared state and memory. A request can plan, call tools, hand work between agents and loop back when a step fails, instead of running a brittle fixed chain.
  2. 2Retrieval that holds up. End-to-end RAG: ingestion, chunking and metadata enrichment, embeddings and vector indexing, then retrieval and context construction across Pinecone, FAISS, ChromaDB and Azure AI Search.
  3. 3Tools the agent can safely use. Tool and function calling into approved APIs, databases and enterprise services through controlled execution, with structured outputs from Azure OpenAI and AWS Bedrock.
  4. 4Evaluation and monitoring. Evals for retrieval quality, groundedness, hallucination risk and latency, plus logging, monitoring and exception handling so problems surface before users notice them.
  5. 5Shipping it. Python REST APIs and microservice integrations, deployed with Docker, Kubernetes and CI/CD across development, test and production.
retryRequestuser or appPlannerLangGraphRetrievervector searchToolsapproved APIsModelOpenAI, BedrockVerifiergrounded?
Fig. 1 How a request moves through the agent graph. Drafts the verifier can’t ground go back to the planner.

Built with LangGraph, LangChain, Azure OpenAI, AWS Bedrock, Pinecone, FAISS, Azure AI Search, Python, Kubernetes, MLflow

Enterprise AI across banking automation, fraud analytics, risk management and customer engagement, where every AI feature also had to clear a bank’s bar for security and governance.

What I built

  1. 1GPT-powered assistants. Conversational assistants and GPT applications on Azure OpenAI that automate banking workflows and support better decisions.
  2. 2Semantic search. Knowledge retrieval over FAISS, Redis Vector Search and Qdrant, so assistants answer from the bank’s own content.
  3. 3Secure AI services. AI microservices and REST APIs integrated with banking systems, with OAuth 2.0, role-based access control and audit logging built in.
  4. 4MLOps and governance. MLflow and Kubeflow for experiment tracking, model lifecycle, controlled deployment and automated retraining, with observability on performance, usage and compliance.
  5. 5Delivery. Automated deployments across AWS and Azure with Docker, Kubernetes, Jenkins and GitHub Actions.
Requestbanking appAccessOAuth, RBACSearchFAISS, QdrantAssistantAzure OpenAIAnswerto the appAudit loggovernance
Fig. 2 Every request is authorized before it reaches search, and every answer is written to an audit log.

Built with Azure OpenAI, FAISS, Redis, Qdrant, MLflow, Kubeflow, Azure Databricks, Docker, Kubernetes, Jenkins

Machine learning for logistics: forecasting, operational analytics, customer intelligence and process optimization, built on large operational datasets and run in production.

What I built

  1. 1Forecasting and prediction. Classification, regression, clustering and forecasting models in Scikit-learn, tuned and validated against the right metric for each problem.
  2. 2Deep learning and NLP. TensorFlow and PyTorch models for complex prediction and unstructured data, plus NLP for text classification, entity extraction and document categorization.
  3. 3Data and feature pipelines. Spark, PySpark, Pandas and NumPy pipelines that clean, transform and engineer features at scale.
  4. 4From notebook to production. MLflow for tracking and versioning, Python REST APIs to serve models, Docker and Kubernetes to run them, and monitoring to keep them healthy.
feedbackOps dataSparkFeaturesPySparkTrainingTF, PyTorchRegistryMLflowServingREST on K8sMonitoringin production
Fig. 3 The model lifecycle, with monitoring in production feeding the next round of data.

Built with Python, Scikit-learn, TensorFlow, PyTorch, PySpark, MLflow, Docker, Kubernetes, PostgreSQL

Predictive analytics and NLP on healthcare data, turning structured and unstructured data into models that supported business and operational decisions.

What I built

  1. 1Predictive models. Classification, regression and clustering in Python and Scikit-learn to surface patterns and trends in large datasets.
  2. 2NLP on healthcare text. Document categorization, text similarity and information extraction for unstructured content.
  3. 3Rigorous validation. Cross-validation, precision, recall, F1 and ROC-AUC, with hyperparameter tuning for stability and generalization.
  4. 4Serving predictions. Spark for the larger datasets and Flask REST services to expose predictions to enterprise applications.
Health dataSQL, textFeaturescleaningModelssklearn, TFValidationF1, ROC-AUCFlask APIRESTAppsdecisions
Fig. 4 From healthcare data to validated models, served to enterprise applications over REST.

Built with Python, Scikit-learn, TensorFlow, Pandas, NumPy, PySpark, SQL, Flask

Backend and database engineering for telecom business operations: the Java services and SQL underneath day-to-day systems and reporting. The foundation everything since has been built on.

What I built

  1. 1Backend services. Java/J2EE applications with Servlets, JSP, JDBC, Spring and Hibernate.
  2. 2Database engineering. Complex SQL, stored procedures, triggers and functions on Oracle and SQL Server.
  3. 3Data pipelines. SSIS ETL workflows and data warehouse structures feeding business intelligence and reporting.
  4. 4Performance and support. Query and ETL tuning, plus production troubleshooting alongside developers, QA and business analysts.
Java servicesJ2EE, SpringDatabasesOracle, SQL ServerETLSSISWarehouseBI modelsReportsBIBusinessoperations
Fig. 5 Java services and SQL feeding the warehouse and the reports the business ran on.

Built with Java, J2EE, Spring, Hibernate, JDBC, Oracle, SQL Server, SSIS

Ask me

A small version of what I build at work. The agent plans, searches this site, drafts an answer and checks every claim against a source before it replies. If nothing here supports an answer, it says so instead of guessing.

Try one of these

A small agent in your browser answers by quoting this site.

Grounded2.45s

“What does Hemanth build?”

He builds agentic AI and copilot systems at Mastercard1: LangGraph agents that call approved enterprise tools, retrieval over internal knowledge, and evaluations that check every answer is grounded before a user sees it2.

  1. 1Experience: Mastercard
  2. 2Skills: agents and evaluation

He builds agentic AI and copilot systems at Mastercard : LangGraph agents that call approved enterprise tools, retrieval over internal knowledge, and evaluations that check every answer is grounded before a user sees it .

  1. planintent: current work
  2. retrieve1 source: experience/mastercard
  3. reasondrafting from 1 source
  4. verify1 claim has no source
  5. retrieveretry, +1 source: skills/agents
  6. reasonredrafting with the new evidence
  7. verifygrounded, 3 of 3 claims sourced
  8. respondanswer with citations
Fig. 6 What the agent did to answer: plan, search, draft, check the draft against its sources.

Skills

The tools I reach for, grouped by the job they do.

Agents
Orchestration, state and tools
LangGraph, LangChain, Multi-agent workflows, Tool / function calling, Agent memory, MCP servers
Language models
Reasoning and generation
Azure OpenAI, AWS Bedrock, Hugging Face, Prompt & context engineering, Structured outputs
Retrieval
Grounding answers in data
RAG pipelines, Embeddings, Pinecone, FAISS, ChromaDB, Qdrant, Azure AI Search
Evaluation
Trust, measured
LLM & RAG evals, Groundedness, Hallucination checks, Observability, Monitoring
Machine learning
Models and data at scale
PyTorch, TensorFlow, Scikit-learn, NLP, Spark / PySpark, Pandas
Platform
Shipping to production
Python, Java, SQL, MLflow, Kubeflow, Docker, Kubernetes, AWS, Azure, CI/CD

Certifications

  • Microsoft Certified: Azure FundamentalsMicrosoft2023
  • Microsoft Certified: Azure Cloud SecurityMicrosoft2023
  • HackerRank Java (Basic)HackerRank2023
  • Using Python to Interact with the OSCoursera and Google2023

Open to senior AI engineering roles, on-site, hybrid or remote

If you’re building something that has to work, let’s talk.

hemanthreddypolu@gmail.com
LinkedIn
in/hemanth-kiran-reddy-polu
Calendar
Book a 30-minute call
Phone
+1 (840) 877-5892
Resume
Download PDF

Or leave a note