Why OCR Breaks Your RAG (And How ColPali Fixes It)
Why traditional text-only parsers destroy financial balance sheets and charts. Comprehensive comparison between VLM Markdown generation and native visual patch embeddings with ColPali.
I architect autonomous multi-agent pipelines, enterprise LLM evaluation platforms, and high-precision hybrid retrieval engines (RAG). Currently leading AI/ML initiatives at Microsoft shipping to Fortune 500 customers across multi-agent sales automation and security threat classification. Previously built core ML platforms at Swiggy and American Express.
Why traditional text-only parsers destroy financial balance sheets and charts. Comprehensive comparison between VLM Markdown generation and native visual patch embeddings with ColPali.
Moving beyond vibes-based testing. Complete guide to decoupling Retriever ranking (Hit Rate@K, MRR, NDCG@10) from Generator Faithfulness with calibrated LLM-as-a-judge CI/CD gates.
Why pure vector embeddings fail on single-keyword deltas, how Reciprocal Rank Fusion ($k=60$) mathematically reconciles candidate lists, and how cross-encoders eliminate semantic collisions. Includes live interactive simulator!
Moving beyond naive static chunking to stateful, self-evaluating retrieval loops. Features an autonomous document grader, query rewrite agent, and fallback decision trees using LangGraph.
Why monolithic God-prompt LLMs fail under complex workflows. How to architect hierarchical supervisors, peer swarms, and standardize tool execution via the Model Context Protocol (MCP).
Emergent multi-agent bugs: circular ping-pong deadlocks and context drift. Complete guide to OpenTelemetry parent-child trace instrumentation and hard token circuit breakers.
The Trojan Resume exploit in production. How indirect prompt injections weaponize privileged tool execution agents, and how to enforce Bell-LaPadula information flow control and 3-tier capability sandboxing.
When autonomy must halt for irreversible actions. How to build durable state suspension, asynchronous approval queues, and cryptographic HMAC-SHA256 signature validation with LangGraph interrupts.
Autonomous multi-agent interrogation game. Every suspect is an isolated LangGraph node with stateful memory checkpointers, prompt injection defenses, and forensic ground-truth deduction evaluation.
Distributed multi-agent residential operations coordinating tenants, owners, and management via Model Context Protocol (MCP). Implements concurrency-safe mutex locks and automated digital gate passes.
100% local-first Chrome extension using in-browser vector search via transformers.js and IndexedDB to auto-fill custom job application forms in under 3 seconds with zero server egress cost.
Interactive mathematical laboratory dynamically balancing dense vector similarity against sparse BM25 lexicons with live cross-encoder re-ranking.
Architecting autonomous multi-agent systems (15M+ msgs/wk across 200+ enterprise tenants) and automated SOC threat triage. Built reusable AgentEval framework reducing validation cycles by 60%, and elevated hybrid RAG retrieval NDCG@5 by 18%.
Maintained core ML platform powering 5,000+ feature jobs and ~100 models in production. Automated model failure turnaround from 1 day to <2 hours, cut AWS cluster costs by 25%, and engineered multi-headed TensorFlow distributed inference.
Engineered 15TB distributed data pipelines using Spark GraphX, Bloom filters, and Levenshtein clustering for merchant entity resolution, speeding up execution time by 30%.
Built 100+ GB network telemetry extraction pipelines for British Telecom and PepsiCo global infrastructure. Developed serverless cloud functions on AWS Lambda (Node.js) for voice AI skills.
🎓 Education: Post Graduate Diploma in Data Science — IIIT Bangalore (2018–2019) • B.Tech in Electrical & Electronics — Dayanand Sagar University (2013–2017)
📜 Certifications: Google Cloud Professional Data Engineer • Databricks Certified Associate Developer for Apache Spark (Scala) • DeepLearning.AI NLP & TensorFlow Specializations