prasnna@github ~
whoami

Prasnna Parthasarathy

Senior Software Engineer

15 years engineering
5+ years on AWS
3 agents in prod

Java and Spring Boot for most of 15 years. AWS for the last five. Building Gen AI agents and MCP tooling in production now. Senior experience + agentic coding tools = a combination that compounds.

scroll

15 years building things

TCS Dec 2009 – Feb 2014

IT Analyst

Full-stack developer on an OLAP analytics platform for AC Nielsen. Java backend, front-end UI, data analysis support. Where I learned that the boring infrastructure work is actually the important work.

JavaHTML/CSS/JSSQLOLAP
Cognizant Feb 2014 – Oct 2017

Senior Software Consultant

Enterprise Java across regulated industries — State Street (regulatory reporting, REST services, in-memory cache for large datasets), NY Life (ACCORD XML transformation via SOAP/Apache CXF), MetLife (batch-to-realtime migration).

JavaSpringRESTSOAPApache CXF
Plymouth Rock Assurance Oct 2017 – Present

Senior Software Engineer

Started building core policy and agency management systems in Java/Spring Boot on Kubernetes. For the last couple of years, shifted to Gen AI: LangChain agents, Bedrock pipelines, MCP tooling — all in production.

JavaSpring BootAWSK8sLangChainBedrockMCP

Stack

Backend
Java / J2EE Spring Boot Spring MVC Python Node.js Kafka
Cloud & Infra
AWS Lambda Bedrock Step Functions Kubernetes Docker
AI / Gen AI
LangChain MCP Bedrock Agent Core Qdrant RAG ReAct
Data & Other
SQL / PL-SQL MongoDB DocumentDB Angular Pandas / Plotly

Things I've built

Plymouth Rock

Claims Processing Pipeline

Multimodal LLM pipeline to analyze claim documents + notes and suggest next actions. Step Functions orchestrates Lambdas — Textract ingests docs, Bedrock does the analysis. The hard part: giving adjusters enough auditability to trust the output.

Step FunctionsTextractBedrockLambdaPython
Plymouth Rock

Policy Chat Agent

Customer-facing agentic chat querying policy, claims, and billing GraphQL APIs plus a RAG store with policy docs. Per-user memory across sessions. Tricky balance: grounded enough to be safe, flexible enough to be useful.

LangChainGraphQLBedrock Agent CoreQdrant
Plymouth Rock

Policy & Agency Management

Core insurance platform — policy lifecycle and agency management. Java/Spring Boot backend on Kubernetes with DocumentDB and Lambda + API Gateway for the service layer. The majority of my time at Plymouth Rock before the AI shift.

JavaSpring BootKubernetesDocumentDBLambda
Side project

Market Rotation Tracker

RRG methodology built from scratch — RS-Ratio and RS-Momentum calculations, live market data, interactive Plotly charts. Started as a "let me understand this properly" project and became something I actually use for my own portfolio decisions.

PythonPandasPlotly
View Analysis →
Cognizant / State Street

Regulatory Reporting Platform

REST service for high-volume regulatory datasets with an in-memory cache layer for throughput. Custom validation framework automating data quality checks. Also: ACCORD XML transformation (NY Life) and batch-to-realtime migration (MetLife).

JavaSpringRESTSOAP / Apache CXF

How I think about this stuff

01

The model is rarely the bottleneck

Most production AI problems are data quality, latency, and trust — not model capability. I've spent more time building output validators and audit trails than tuning prompts.

02

Start with low autonomy

The prod support agent starts with workflow tools and human-in-the-loop steps. Agents that do too much too fast are hard to debug and harder to get sign-off on in regulated environments.

03

Enterprise background pays off in AI work

15 years in regulated environments means auditability, failure modes, and compliance are first-class concerns — not afterthoughts you bolt on when something goes wrong in prod.

04

Skeptical of the hype cycle

Agents are genuinely useful. They're also fragile in ways demos don't show. I've shipped both the wins and the failure cases — the second teaches you more.

05

Agentic coding is a force multiplier — if you know what correct looks like

I use AI coding assistants heavily and the velocity gain is real. What makes it work is 15 years of knowing the right architecture, the edge cases, the abstractions that will bite you in six months. Without that, you're just generating code faster. With it, you're compounding.

Get in touch

Open to conversations about enterprise AI tooling, agentic systems, or anything backend. Not actively looking but happy to talk.