Anup Das

About Anup Das

Your AI demo is lying to you.It runs on cleaned data, patient users, and prayers. Real production has none of those.I know because I\'ve built both kinds. THE MOMENT EVERYTHING CHANGED2019. I shipped an NLP system that analyzed customer feedback.In testing: 98% accuracy. Instant responses. Executive team loved it.In production: Response times hit 45 seconds. The system choked on real customer slang, emoji, and typos we never tested. Within 72 hours, support tickets tripled.My manager asked one question:\"Did you test this with actual users?\"I hadn\'t.That failure rewired my brain:→ Demos optimize for applause→ Production optimizes for survival → The gap between the two will destroy youSo I spent the next 7 years obsessively learning what separates AI that ships from AI that dies in staging. WHAT I BUILD NOWToday at EY, I lead data scientists building GenAI systems with one non-negotiable rule:If it can\'t survive production chaos, we don\'t ship it.The results:→ 1M+ documents processed daily (zero downtime in 8 months)→ <1s response times under real enterprise load → Multi-agent architectures that scale, not crashI work with Custome Agents, LangChain, CrewAI, and Model Context Protocol (MCP) not because they\'re trendy, but because they solve problems demos ignore. WHAT I TEACH (So You Don\'t Learn the Hard Way)I document every production failure and fix on Medium (@anuptechtips) for 769+ engineers who are done with AI hype:• Multi-agent systems that survive real users• LLM optimization beyond \"just add more GPUs\"• GenAI architectures that handle enterprise chaos• The mistakes that killed my demos (and how to avoid them)Not theory. Not tutorials. Just battle scars from shipping AI that works. WHO THIS IS FORYou\'re building AI and:→ Your demo is impressive but you\'re terrified of production→ Your system works great until real users show up → You\'ve heard \"it worked in testing\" too many times→ You\'re tired of AI content that ignores the hard partsSound familiar? Let\'s connect. I\'ve crashed enough systems for both of us. LET\'S CONNECTBuilding AI that needs to survive production? Working on multi-agent architectures?→ Read my production lessons: https://anuptechtips.com→ DM me: I respond to every message about production AI→ Open to remote opportunities building systems that actually scaleFair warning: I\'ll ask if you\'ve tested with real users yet.

Experience

1. Associate Manager

   EY

   Jan 2024 — Present · Kolkata, IN

   Led a team of 6 data scientists and ML engineers, delivering Generative AI, LLM, MCP server and agentic AI solutions for enterprise-scale clients.• Architected a Generative AI platform to auto-generate Terraform & CloudFormation templates with built-in policy checks, cutting manual scripting by 70% and reducing deployment errors by 60%.• Designed ingestion pipelines handling 1M+ documents (Confluence, GitHub, PDFs, PPTs), enabling sub-second semantic search with FAISS, Pinecone, and PostgreSQL.• Developed a Model Context Protocol (MCP) server exposing cloud APIs to LLMs, orchestrating autonomous agents for real-time data analysis and code generation.• Tuned CUDA/NCCL for GPU-accelerated training and inference on AWS/GCP Kubernetes clusters, achieving 3× throughput improvement.• Built an asynchronous embedding pipeline using AWS Lambda, SQS, and SNS to process 500K+ documents daily, improving reliability by 40% and reducing latency by 25%.

Education

• Dr. B.C. Roy Engineering College

  Master’s Degree, MCA

  2015 — 2018

• Siliguri College

  Bachelor's Degree, BCA

  2012 — 2015