I design and ship agentic AI systems — from foundation-model training to production LLM agents and retrieval.

Pretraining a large-scale language model on a ~15–20 trillion token corpus spanning Indic languages and Indian domain knowledge. Owns the full pipeline — data curation, tokenizer design, and training pipeline design. Builds on prior hands-on experience training and fine-tuning smaller language models, applying that foundation-model background to grounding, evaluation, and failure-mode analysis in downstream agent and RAG systems.

A multi-LLM developer support agent (Claude, GPT, Gemini) on FastAPI + Next.js with RAG, FAISS semantic code search (tree-sitter indexing), and incident debugging through a unified platform. Implemented guardrails with human-in-the-loop approvals, complete audit trails for compliance, and automated diagnosis for API latency issues and codebase knowledge discovery.

An intelligent phone-automation platform combining Vapi voice infrastructure, OpenAI GPT-4o, and Qdrant vector DB for natural, real-time customer conversations with direct Zammad ticketing. Real-time knowledge-base retrieval via vector embeddings during live calls; a post-call GPT-4o pipeline extracts sentiment/intent and auto-creates enriched tickets with AI-generated tags. Designed a 7-collection Qdrant architecture with PII redaction, HMAC webhook auth, input sanitization, and SMS/email follow-up scheduling.

A Chrome extension with autonomous agentic capabilities — deep analysis, web search, email composition, save/bookmark — using Gemini 3072-dimensional embeddings for semantic page indexing and RAG-based retrieval. Enables cross-session browsing memory with intelligent section highlighting, context-aware search, adaptive theming, and export (PDF, text, clipboard).

End-to-end Bronze→Silver→Gold pipeline processing 7 SACWIS tables (10M+ records). Silver layer with SHA-256 SSN hashing, dedup flags, and stored-procedure cleaning; Gold MERGE upserts with surrogate keys and a 29-entry audit log.

Developed a Convolutional Neural Network (CNN) for breast cancer histopathology image classification using ~30,000 histology images, achieving ~93% classification accuracy with a 25% inference optimization, deployed as a REST API.