Headline

Scaling AI Agents in Production | AI Engineer | LangChain · LangGraph · RAG · MCP · Amazon Bedrock | Python · Java · React Next.js

About

I build AI systems that actually ship. Not demos — things people use every day. At Nomura, I found a validation problem everyone had accepted as unfixable: a financial compliance report with 150 columns × 60 rows that took about a week to validate manually. I pitched it, designed it, and built an AI web validation agent — LangChain, Playwright MCP, in-house LLM APIs, with deterministic tool-calling flows so it can't hallucinate its way through financial data. It now does the job in ~7 minutes and developers use it daily to self-test before QA handoff. I owned the whole thing: spotting the problem, getting buy-in, building it, driving adoption. Same with the test automation framework I built there from scratch — cut the regression pack from 80 minutes to 55. Other things I've built: A RAG resume screening system running as a distributed deployment — Celery workers on Kubernetes, Amazon Bedrock for inference, FAISS retrieval with cross-encoder re-ranking, RAGAS for evaluation, LangSmith for tracing. A government-scheme RAG chatbot (FAISS + Neo4j hybrid, 23 Indian languages) — Top 105 globally at GDG Solutions Challenge 2025, out of 64,000+ registrations. Also: an e-commerce microservices platform on AWS EKS, and a real-time chat system handling 500 concurrent users with an 80% latency improvement. Stack: Python, Java, TypeScript · LangChain, LangGraph, CrewAI, AutoGen · RAG, FAISS, Pinecone · MCP (FastMCP, custom servers) · Anthropic API, Amazon Bedrock · Docker, Kubernetes, AWS · React, Next.js, Spring Boot LeetCode Knight — 1800+, top ~3% globally. Anthropic-certified: MCP Advanced Topics, Claude with Amazon Bedrock, AI Fluency. B.Tech in Computer Science (AI & ML), SPIT Mumbai, June 2026 graduate. Open to AI Engineer roles. prathamshirbhate1909@gmail.com

Location

Mumbai

AI stack

claude code

codex

cursor

github copilot

mcp

Work experience

SDE 1 at NxtPe, Jul 2026 - Present · 1 mo, remote, AI Microservice & RAG : • Built a centralized AI routing microservice and agentic workflows for production LLM traffic — prompt/task-based model routing, image/Base64 processing — with observability into token usage, model parameters, confidence, and routing behavior for measurable platform impact. • Built a production RAG pipeline over a vector store, converting unstructured business knowledge (PPT/PPTX) into metadata-rich, incrementally updatable embeddings; paired with a recurring weekly/biweekly evaluation loop that uses observed production behavior to improve agent routing and triage logic. Chattyv2 : • Built a configurable conversational commerce engine (Java, Spring Boot) shipped to production; designed a 5-stage deterministic message resolution pipeline (interactive reply → alias → regex → configurable LLM intent-triage fallback → human escalation), sustaining 20k messages/sec peak throughput. • Engineered a resilient, resumable hot-path / cold-path architecture over an Apache Kafka broker with concurrent, retry safe asynchronous execution: the hot path resolves messages against compiled flows in a Caffeine cache backed by Redis (off PostgreSQL) while the cold path asynchronously persists execution records — decomposed into 5 sub-modules within a single pod-scalable microservice on Kubernetes. • Built a durable, graph-based workflow engine modeling conversational journeys as executable nodes/transitions with bounded state and breadcrumb traversal — enabling deterministic, resumable and fully configurable execution that can resume from controlled states.

Software Engineer Intern at Nomura, Jan 2026 - Jun 2026 · 6 mos, Mumbai, Compressed a week-long financial compliance report validation (150 cols × 60 rows) to ~7 minutes by building a self-initiated AI web validation agent using LangChain, Playwright MCP, and in-house LLM APIs, with deterministic tool-calling flows as hallucination guardrails — now in daily production use by developers for pre-QA self-testing. Owned end to end: problem discovery, design, build, and adoption. Cut repetitive 4-minute test flows to under 15 seconds by engineering custom MCP tools; the agent auto-generates verified Python Playwright regression scripts and files Jira defects autonomously, reducing early-stage defect leakage. Reduced regression pack execution 31% (80 → 55 min) by architecting a UI automation framework from scratch in Core Java, Selenium WebDriver, and Cucumber-Gherkin BDD for a globally deployed equities trading platform; delivered 2 full release sign-offs in a single week.

Full Stack Developer Intern at Octosell AI, Oct 2024 - Jan 2025 · 4 mos, Remote, Expanded the paid-tier feature surface of a production Shopify plugin by shipping 2 post-subscription features (analytics chat logs panel, setup guide panel) using React, TypeScript, GraphQL, and Node.js. Eliminated settings drift across 3 surfaces (admin panel, preview editor, theme editor) by building a metafields-based sync system. Reduced client-reported UI inconsistencies by overhauling mobile and desktop plugin UI/UX into reusable, responsive React components.

Education

Bachelor of Technology, Computer Science (AIML), Bhartiya Vidya Bhavans Sardar Patel Institute of Technology Munshi Nagar Andheri Mumbai

Public links

github: https://github.com/pra2107tham