mehul
srivastava

I build tools that have to live with people. Right now that is Marshmallow, an open-source memory runtime for agents, and the sourcing and evaluation systems a venture fund runs on at Biome. Before that: lead engineer on an LLM guardrailing platform at HCLTech, the one person on product full time at a social network as it grew to 2.5 million users, and a college social network I started that was acqui-hired.

The question underneath all of it: auditable personal memory for agents, interfaces that make multi-agent work inspectable, and generated environments that keep a human-owned record. The essays live at probably semicolons. The application details are in the margin and on the cv.

How should intelligent agents, and teams of agents, remember, reason, and remain legible to the people using them?

Three attempts, one per word. Memory: Marshmallow, so an agent does not walk in cold. Reasoning: Indra, a design study for an operating system for teams of agents, self-published as a whitepaper, which shrank into Marshmallow once the colony turned out to cost more than anyone would pay. Legibility: Biome, where a ranking the partners could not see the reasons for was a ranking they stopped using.

The years before those, Slick and Circles, were the startup years. They are on the site as records, with what I took from them, because the question started there. The route is in the story.

The two-page CV is here, also as a pdf. If you run a lab, a programme, or a team working on agents and the people who use them, write to probablymehul [at] gmail [dot] com, or find me on LinkedIn and GitHub.