2026-09-23 –, Main Stage All times in US/Mountain
We built Bond, an internal AI platform that helps our engineers with documentation, incident response and CI/CD investigation. To do that well it builds a temporal knowledge graph of who actually answers what, how confident that signal is and how it changes over time. We built it to improve Bond's responses; it produced something we didn't plan for and revealed a preference map of how engineering knowledge actually flows in the organization which (un)surprisingly looks nothing like our org chart. This talk is about the gap between the two, why it matters and what every internal AI platform is already telling you about your team if you look.
Most engineering organizations are sitting on a dataset they've never looked at: the trace of how knowledge actually moves between their engineers. We built Bond which is an internal AI platform at Klaviyo to assist with documentation, incident response, CI/CD investigations, and project management. It touches the practical surface area of the software delivery lifecycle which is to say, the entire organization. To do that well Bond needed a memory. We gave it complete CRUD access to a temporal knowledge graph database backed by AWS Neptune and structured around Graphiti that grows with every interaction. A year in, that graph has almost become more interesting than the platform it was built to serve.
Bond is a multi modal system where engineers reach it through Slack, APIs, webhooks, and scheduled tasks. Behind it sits many specialized subagents covering documentation, platform and observability, project management and other workflow surfaces. Every conversation Bond participates in gets classified with who asked, who answered, who corrected the answer, who routed someone to the right expert and in what slack channels that was done in. Expertise is modeled as confidence scores against domains, strengthened by demonstrated knowledge and decayed over time. The result is a participant graph that captures not just what people know, but how that knowledge moves, who connects it, and how it shifts as teams and technologies change underneath them.
I am framing the talk from behavioral economics. Stated preferences are what people say they want; revealed preferences are what their actions show they actually want. An org chart is a stated preference for where knowledge lives and over time, Bond's graph has shown the true revealed preferences. I'll walk through three patterns the graph surfaced that the org chart didn't: hidden experts (the person everyone actually asks, who doesn't own the system on paper), invisible routers (people whose value is connecting others to knowledge rather than holding it themselves) and expertise shifts (confidence decay revealing organizational change before the org chart catches up).
As well as why this matters operationally. The same graph that reveals organizational structure is what lets Bond route incidents to an actual expert rather than the official owner, diagnose flaky tests by connecting them to recent deployments and config changes across systems and accelerate onboarding by pointing new engineers at both the right person and the right document.
I'm a Platform Engineer at Klaviyo, where I've spent the last four years building developer infrastructure across site reliability, DevOps, and AI enablement as the team has reorganized. My current focus is Bond, Klaviyo's internal AI platform and STEP our ephemeral test environment system as well as the many other CI/deployment systems we have always maintained. I'm a mechanical engineer by training (Stevens Institute of Technology, NJ) who took the long way into software via SRE. I live in Colorado Springs with my wife our cat Octavia our golden retriever Denali, and are avid local hiker/camper/skiers and am mid-conversion of a Ford Transit into a camper van (just CO things).