Leo Mehr

Leo Mehr co-founded Lumos and leads Lumos Labs, its applied-AI initiative focused on making autonomous agents safe for enterprise use. His work addresses a practical security problem: how to give agents access to sensitive systems without losing control over their permissions, decisions, or actions.

Mehr studied engineering physics at Cornell, spent a year studying computer science and philosophy at Oxford, interned at Google, worked at Hudson River Trading, and studied machine learning and systems at Stanford. At Stanford, he met Andrej Safundzic and Alan Flores-López, his future Lumos co-founders.

Lumos launched publicly in 2022 with more than $30 million in funding, including backing from Andreessen Horowitz and Neo. Its original product gave employees a self-service way to obtain software access while helping IT and security teams manage permissions, approvals, and licenses. Mehr helped build the product and led an engineering organization of roughly 25 people before joining Ramp.

At Ramp, he became an engineering director responsible for forward-deployed engineering, developer APIs, AI services, and the agent platform. His team launched Ramp’s MCP server, bringing agents into financial workflows and exposing the limitations of access controls designed for human users. His approach to forward-deployed engineering places customer-facing engineers inside the engineering organization, where they can solve enterprise problems without compromising product coherence.

• Customer scoping determines product quality. An urgent SAP integration request calls for examining the actual deadline, existing workarounds, customer API capabilities, and whether other customers would benefit. Ramp engineers once built a reimbursement feature for iOS and Android before discovering that the customer issued only iOS devices—a costly lesson in validating basic assumptions.

• Agent-assisted engineering requires human judgment. At Ramp, Slack requests flowed into Notion, where an agent asked follow-up questions and helped turn incomplete submissions into specifications. Mehr sees opportunities to automate context gathering, scoping, specifications, and implementation, but emphasizes reliable evaluations, product context, intermediate orchestration, and human accountability for the result. His AI Engineer conference talk develops these ideas through concrete enterprise examples.

• Agent identity needs enforceable boundaries. Autonomous agents can operate continuously across multiple systems, accumulating permissions and executing sensitive actions faster than people. Mehr advocates short-lived, contextual authority, externally enforced policies, audit trails, human approval for consequential actions, and mechanisms for containment and reversal.

In August 2026, Mehr returned to Lumos to launch and lead Lumos Labs, focusing on agent governance and access management for enterprise workflows involving money, production infrastructure, and sensitive customer data.