I make software, and AI, work where the work happens.
I'm Vamsi, a senior software engineer with 12+ years of building production software. I spent nine of those years in Java: Spring Boot microservices, REST and SOAP integrations, Kafka and the databases underneath, for companies where downtime and bad data weren't options.
Since 2023 I've been building AI systems on top of that foundation: LLM, RAG and multi-agent applications for enterprise teams. I build them end to end, from the React front end to the FastAPI services, the data and retrieval pipelines, the AI layer and the AWS infrastructure they run on.
My favorite part of the job is the part before the code: sitting with the people who do the work, finding out what actually slows them down, and agreeing on what “good” looks like before anyone writes a line.
The short version
- I own systems end to end: architecture, APIs, data, front end, AI, evaluation, deployment and handoff.
- My base is backend and distributed systems. The AI work is built on it, not instead of it.
- I care more about “is it right, and can we prove it?” than “is it clever?”
- I'm open to Forward Deployed Engineer and senior AI engineering roles, where the job is to sit close to a customer and ship something that works for them.
What I build with
- Front end
- React, TypeScript, JavaScript, Streamlit
- Backend and distributed systems
- Java 8/11, Spring Boot, Spring MVC, Python, FastAPI, REST, GraphQL, SOAP, microservices, Kafka, OAuth 2.0, mTLS
- Data
- Aurora PostgreSQL, Oracle, DynamoDB, pgVector, FAISS, Pinecone, ETL and embedding pipelines
- AI
- Claude, GPT-4o, AWS Bedrock, Azure OpenAI, LangChain, LangGraph, CrewAI, RAG, multi-agent orchestration, MCP, guardrails, RAGAS evaluation
- Cloud and DevOps
- AWS (Lambda, Step Functions, API Gateway, EKS, S3, IAM/KMS), Docker, Kubernetes, Terraform, GitHub Actions, Jenkins, CloudWatch
How I got here
- 2014
Napier · Java developer: SOA integrations and REST services for telecom and healthcare clients.
- 2015–2021
Client: Apple · Spring MVC and Spring Boot applications for product data management.
- 2021–2022
Wells Fargo · Spring Boot microservices for financial change management.
- 2022–2023
TIAA · Spring Boot microservices on an enterprise platform, with zero production incidents.
- 2023–now
Production AI · Building LLM, RAG and multi-agent systems end to end on AWS. The model is the easy part; the system around it, and trust, are the hard part.
What I believe about building software and AI
AI is one layer, not the whole system.
Most of what makes an AI product work is ordinary good engineering: clean APIs, reliable data, retries and timeouts, security, monitoring. Nine years of Java taught me that before AI did.
Measure before you claim.
If I can't show the eval, I don't quote the number. When my own metric was too generous, I made it stricter and published the lower score.
“I don't know” is a feature.
A good assistant refuses or asks when the documents don't say. A confident wrong answer costs more than no answer.
Humans stay in the loop where it matters.
Low-confidence outputs go to a person instead of a guess. That one rule held 90%+ field accuracy in production.
Build it so they can run it without me.
Clear docs, tests, cost limits and monitoring. A handoff is part of the product, not an afterthought.
Why a hotel project
Hotels run on policies scattered across binders, emails and people's memory, and front-desk staff need answers in seconds, with a guest standing in front of them. That makes it an honest test of what an AI assistant must do: answer from the right property's documents, cite the source, and say “check with a manager” when it isn't sure.
It's an independent project using a simulated hotel group and synthetic data. I'm building it like a real client engagement in five stages, from the first assistant to the final handoff, and publishing each stage with its real numbers, tests and trade-offs.
How I work
- 1
Discover: sit with the people who do the work; write down the real problem and what “good” means.
- 2
Measure first: build a small evaluation set before building features, so every change is judged by numbers.
- 3
Ship small: short iterations, each one tested, evaluated and shown to users.
- 4
Secure by default: secrets never in code, input and output guardrails, cost limits, least-privilege access.
- 5
Hand off cleanly: docs, runbooks, tests and monitoring so the client's team can run it without me.