Deepak Mittal

I’m an Applied AI Engineer focused on fine-tuning and optimizing state-of-the-art LLMs for real production constraints. At my current role, I’ve fine-tuned Llama 3, Mistral, and Gemma using Unsloth, HuggingFace, and PEFT (LoRA/QLoRA) with 4-bit quantization, then evaluated and improved performance with Promptfoo.

I build end-to-end GenAI products, including a multimodal Product Discovery Engine that processes 2.9M+ product images. I used Gemini multimodal for text attribute extraction and dual text+image embedding search backed by Elasticsearch, orchestrating BigQuery sync, embedding generation, and knowledge indexing with Vertex AI Pipelines and Kubernetes Jobs.

I also architect agentic AI systems and RAG pipelines that teams can actually operate. Using LangChain, CrewAI, and Google ADK (with vector databases like Chroma, FAISS, and Elasticsearch), I built multi-agent orchestration with MCP Server, A2A protocol, and autonomous tool calling via Celery async queues—reducing workflow execution time by 65%.

Previously, I worked as a GenAI Data Scientist, leading an enterprise QA bot ingesting 12,000+ product PDFs and improving inquiry resolution by 40%. Before that, as a Data Scientist in Product & AI Research, I directed R&D that boosted classification accuracy by 20% and improved information extraction accuracy by 26% with LayoutLMv2, while also helping develop IDP Studio—a scalable document processing platform built with Python, microservices, and Kubernetes CI/CD.