Pellametric: Seeing Inside the Token Black Box

For the past couple of weeks, I joined forces with Walid Khori, Gerardo "Sebastian" Garces, David Aihe, and Sandesh Pathak to address a fascinating problem that we identified in the AI industry for our Gauntlet Capstone project.

Token spend is now higher than ever, but clear visibility around usage, best practices and efficiency hides behind the black box that is model providers.

Sure, AI model providers have completely redefined the way that we do work, but their incentives are clearer today now than ever: use as many tokens as possible, as fast as possible, the more the better.

In order to address this, we decided to build an open-source analytics dashboard aimed at empowering engineers and engineering teams with the insights necessary to improve their usage, identify trends and patterns, and recognize what's working and what isn't.

Pellametric pulls your machine logs from across all your AI providers, synthesizes the data and provides you and your team with a single unified dashboard that gives you visibility into where your token spend is going

How many tokens did you spend on your last PR, how many tokens were spent using a given Claude Code skill or a given MCP.

Understanding what's going on inside the black box will key for the success of many teams in the AI race. We're living in an unprecedented time where token usage is being hyper-subsidized, but I'm convinced it's not going to last forever.

---

Originally posted on LinkedIn.