Headline

Fixed a production LLM memory bug that cut latency 150x. Built multi-agent AI for a global bank. | LangGraph · Google ADK · FastAPI

About

Most engineers learn to code. I learned to read muscle signals first. My first real engineering project was a myoelectric prosthetic arm. EMG sensors on the residual limb. An Arduino translating electrical signals into servo motor commands. A mechanical hand that responded to human intent in real time. No framework, no tutorial. You figure out why the signal drops at 3am, or the hand doesn't move. That's where I built my relationship with latency. The gap between intent and response. Hardware teaches you that gap is a problem that doesn't get to stay. Before I wrote production AI, I spent 8 months on the phone — 20,000 conversations with CFOs, VPs of Finance, and Treasury leads across LATAM, EMEA, and APAC, diagnosing where their O2C and invoice workflows actually collapsed. Not in a case study. In live enterprise environments, under pressure, where a failed process run costs real money. Most AI engineers have never seen that. I've seen it 20,000 times. Those two things — hardware precision and enterprise chaos — are what I brought into software. At Capgemini, I'm building a production multi-agent AI platform for a global banking client. Parallel LLM agents handle content generation, image creation, compliance validation, and multi-channel delivery — all triggered by live Adobe Analytics signals. Marketing teams set the segments. The agents execute everything else. Result: ~30% less manual campaign work at enterprise banking scale, zero corruption across concurrent agent workflows. In parallel, I am contributed to a production LLM memory system at voice AI startup — took agent memory retrieval from 150ms to 1ms. Not a benchmark. A live system with real users depending on it. Grand Finalist — TVS Motor Innovation Challenge 2024, IIT Madras, 500+ teams competing. Top 50 — Tata Imagination Challenge 2024. I build agentic AI for people who need it to work in production — not in a demo, not as a chatbot wrapper, in a real system making real decisions when no one is watching. Stack: Python · LangGraph · Google ADK · FastAPI · GraphQL · pgvector · React.js · AWS · Docker If you're building something where agents need to make real decisions in a real environment — not generate text, actually decide — reach out directly: shreyan826905@gmail.com

Location

Bengaluru

AI stack

claude code

langchain

langgraph

mcp

huggingface

Google ADK

Work experience

Software Engineer at Capgemini, Sep 2025 - Present · 1 yr 1 mo, Bengaluru, Python · FastAPI · GraphQL · LangChain · LangGraph · Google ADK · Gemini APIs · PostgreSQL · Adobe Analytics · AWS (S3, EC2, SQS) · Agentic AI · Multi-Agent Systems · Prompt Engineering · REST APIs · Microservices · CI/CD · Agile · TDD · Unit Testing · Integration Testing Eliminated ~30% of manual campaign setup at enterprise banking scale — engineered a production multi-agent platform (Google ADK + LangGraph + LangChain) orchestrating parallel agents across content generation, image creation, compliance validation, and multi-channel delivery, all triggered by real-time Adobe Analytics signals Enabled marketing teams to launch personalised AI-driven campaigns with zero engineering intervention per cycle — architected GraphQL + FastAPI user segmentation APIs connecting real-time Adobe Analytics signals directly to agentic inference workflows Delivered sub-100ms feature retrieval on a live banking system — designed end-to-end data pipelines (Adobe Analytics → PostgreSQL with composite indexing → LLM inference layer) ensuring real-time context was always available for autonomous agent decisions in production Delivered banking-grade data integrity with zero corruption across parallel agent workflows — built RAG-ready PostgreSQL + pgvector infrastructure with multi-database approval/rejection state management handling concurrent live decisions

Operations Associate at AppsForBharat, May 2025 - Sep 2025 · 5 mos, Bengaluru, Karnataka, India, Python · SQL · n8n · Firebase Studio · Adobe Analytics · Cursor · AI-assisted Automation · Cross-functional Ops · Stakeholder Management · Process Improvement · Workflow Automation Cut operational turnaround ~40% and improved data accuracy 20% — built Python + n8n AI-assisted automation pipelines replacing recurring manual processes across 1,000+ live puja workflows; recognised directly by the Founder for engineering impact Owned end-to-end fulfilment operations at a 100,000+ user religious-tech platform — coordinated 6+ cross-functional teams across scheduling, logistics, and delivery for 200+ temple and priest partners nationwide where platform trust was the product

ABM & Fintech Domain Analyst at HighRadius, Jun 2024 - Jan 2025 · 8 mos, Bhubaneswar, Odisha, India, Developed the enterprise operational context that now informs how I design agentic AI systems — 20,000+ direct conversations with CFOs, VPs of Finance, and Treasury leads across LATAM, EMEA, and APAC diagnosing broken O2C, invoice automation, and treasury workflows at scale Built deep domain fluency across O2C, R2R, P2P, and Treasury processes — the operational knowledge that separates AI engineers who understand enterprise chaos from those who only understand APIs

Junior Developer at Droupout, Jun 2024 - Aug 2024 · 3 mos, Remote, Python · React.js · AWS (S3, EC2) · MySQL · REST APIs · WebRTC · HLS · JavaScript · Bootstrap · Git · Full-Stack Development · SaaS · Real-time Systems · Live Streaming · Backend Engineering Shipped a full EdTech SaaS platform from zero infrastructure — MySQL schema design, Python REST APIs, React.js frontend with WebRTC live streaming — serving 10,000+ active learners on AWS (S3 + EC2) Reduced live video streaming latency 20% under concurrent load and resolved 10+ production incidents through systematic root-cause analysis — this was the role that taught me I watch production, I don't just ship and move on

Industrial Trainee at Steel Authority of India Limited, May 2023 - Jun 2023 · 2 mos, Burnpur, West Bengal, India, Industrial Operations · Manufacturing Systems · Process Engineering 2-month industrial immersion at one of India's largest steel producers — Burnpur (IISCO Steel Plant): • Observed and documented end-to-end steel production workflows — blast furnace operations, rolling mills, quality control, and plant logistics management. • Gained hands-on understanding of industrial control systems, plant automation, and large-scale manufacturing operations — strengthening hardware-software integration perspective. • Developed understanding of industrial process engineering and operational management at national infrastructure scale.

Education

Bachelor of Technology - BTech, Electrical, Electronics and Communications Engineering, KIIT - Kalinga Institute of Industrial Technology

10 +2, Mary immaculate school

Public links

github: https://github.com/mustang-shr

leetcode: https://leetcode.com/HridxXxXxX