The engineering around the model is where the product happens.

I started in enterprise software and spent over a decade in serious, production-grade codebases: an MDM platform at VMware serving large device fleets, C# reporting systems, and three years in Workday Financials shipping full-stack features in invoicing and a resume-processing module used by thousands of enterprise customers. Workday is also where I first built AI into a product and where I stepped into product ownership, so I learned to work backward from the customer, not just the ticket.

Two years ago I went out on my own to build AI products from scratch. Ganit is an agent that replaced a manual back-office job at a wholesale distributor: staff used to look up dealer-specific pricing in spreadsheets during phone calls, and the agent now does it instantly over Telegram. The hard part was making a non-deterministic model trustworthy enough to quote real prices, so I built hybrid retrieval plus an LLM reranker with a disambiguation step, and pushed anything that had to be exact (like the actual price math) out of the model's hands entirely. Lead Surf is a LinkedIn lead-intelligence Chrome extension that reads a user's feed and classifies buying signals, where the interesting problems were cost control, running third-party text through an LLM safely, and calling LinkedIn's own API from the browser without tripping its abuse detection.

The thread through all of it: I like owning the whole thing, discovery through production, and the engineering I find most interesting is the product layer that makes an unreliable model reliable enough to trust.