Hemant Virmani

• First Principles: Trained transformer-based LLMs from scratch; studied attention mechanisms, embeddings, and fine-tuning vs. production trade-offs.
• Agentic AI: Built a multi-agent GAIA benchmark system with LangGraph, scoring well above targets on Huggingface's GAIA benchmark.
• DeskGenie (open source): Desktop automation agent — LangGraph + FastAPI + React, multi-backend. Scored 100% on Kaggle's SAE benchmark.
• ZoneBridge: Agentic systems project — because building is a habit, not a phase.
• Thought Leadership: Engaging with technology leaders on AI-driven engineering, org design, and GenAI transformation across industry, especially in logistics and eCommerce.

At Amazon's Bellevue, WA office, I led Engineering and Science for Customs and Compliance across Amazon Global Logistics. The question underneath all of it: how do you make sure every package crossing a border plays by the rules — whichever country's rules those happen to be? Over 4.5 years: expanded from US-only to UK, EU, and Japan, across three Amazon businesses; built a science team from scratch in 2024; shipped an agentic AI system that cut classification costs by 95%. That last one still surprises me — it's what happens when you approach an old manual process like you've never seen it before.

At Amazon's Hyderabad, India office, my work evolved over seven years in a way I didn't fully anticipate. It started with infrastructure — a real-time transportation monitoring platform handling 10K+ TPS. Then shifted toward something I found more interesting: computer vision and document extraction for seller verification, including forgery detection — work that later became AWS Textract's AnalyzeID API. In between, I re-architected Amazon's de facto secure document storage system, the kind of foundational work nobody sees but everything depends on. The question I kept coming back to: how do you know the person signing up to sell is actually who they say they are? By the time I was leading in-person seller identity verification end-to-end, all of it had compounded — a system that onboarded 2M+ businesses and cut fraud to a sixth of what it was.

Spent over a decade at Adobe, Noida, building and scaling PS/ PDF/ XPS/ Print-publishing platforms. Served across roles from Member of Technical Staff to Senior Engineering Manager. Led a research team to launch Adobe LeanPrint, co-authored six US patents, shipped a PS/PDF-to-XPS converter in print pipelines, and delivered the long-awaited Multiple Undo feature in Adobe FrameMaker, among other platform and product improvements.

Software Engineer at startup Zenax Technologies (formerly, Cynosure Media Solutions) at Noida, building 0-to-1 products across Linux and web stacks, including NAS virtualization for US partners and a Linux multitasking server messaging framework.

Completed M.S. in Software Systems from BITS Pilani through Distance Education while working full-time at Adobe. It helped me deepen my systems knowledge and satisfy my curiosity by going deeper into how software systems work.

Completed M.Sc. in Information Technology from Punjab Technical University through Distance Education while working full-time to build a stronger IT foundation early in my career.

At Delhi College of Engineering (B.E.), I built an engineering mindset for structured problem-solving. Although I majored in Civil Engineering, I kept exploring software, learned programming in parallel, and used that foundation to build a long-term technology career.