As enterprises shipped LLM features, AI spend became their fastest-growing, least-understood cost line — scattered across provider dashboards with no answer to “which application drove this spike?” or “what does a customer cost to serve?”

I led the product from spec to delivery: unified spend and usage across OpenAI, Gemini, Anthropic and Bedrock; near-real-time anomaly detection on token spend; unit economics down to cost-per-token, cost-per-request and cost-to-serve per user; and attribution mapping every dollar to the applications, teams and users behind it.

A FinOps analyst's day is a loop: someone asks a cost question, the analyst finds the right data, runs the analysis, spots what matters, builds the dashboard. We built AI agents that run that loop from a single natural-language request.

I designed the intent-and-analysis routing (how the agent decides what to analyze and which data to reach for) and the human-in-the-loop review workflow that keeps an analyst in control of what ships. I prototyped the experience hands-on in Claude Code and Figma Make — engineering took it to production from that prototype.

Enterprises couldn't answer “what do we actually run, and how is it configured?” without stitching together three cloud consoles. I led a net-new product — design to production — that unifies resource and configuration data across AWS, Azure and GCP into a single searchable system of record.

Under the hood: a relationship graph of cloud assets with a DSL-based query interface, so customers can discover, govern and secure resources with integrated policy controls.

Inconsistent native cloud tags quietly break cost allocation — nobody can tell which team drove which spend. Virtual Tags lets customers define rule-based labels applied across clouds, unifying messy tags and attributing spend to the right teams and projects with zero infrastructure changes.