The idea for PeopleOS came from a simple need: having a home base at work. Most mornings, I’d log on and have no idea where to start. I imagined a platform that would give me clear benchmarks for my progress, a single place for everything related to benefits and pay, and an agent to connect me with whatever information I needed when I felt blocked or unsure.
My experience in corporate was surprisingly isolating. In a post-COVID world, all interactions happened over Slack. As a junior engineer, there was no osmosis of knowledge—just endless hours blocked, waiting on people who had other priorities, feeling disconnected from my team and unsure what to do next. HR felt invisible, and the recruiter I’d built a relationship with before starting disappeared on day one. I was left to figure things out on my own, relying on whoever happened to be available.
Iteration 1: PeopleOS Platform
I started exploring the idea of personal agents for employees—a digital home base you could return to throughout the workday. When you’re stuck, unsure how to follow up with someone, or searching for the info you need to do valuable work, your assistant would be there to guide you. This iteration was definitely more of a design concept than a real product - later, I drilled down into what the actual problem was, explored technical feasibility, and built a few demos.
Iteration 2: Slackbot Assistant
As I talked to friends, I realized employee experience is a lot about what you know and who you know—the expertise you build, your engagement on Slack, and the meaningful connections you make. I decided to experiment with a Slackbot that understood the company’s entire workspace, embedded conversations in a knowledge graph, and could answer questions about who knows what, suggest channels to join, and help with more intelligent searching when you don’t know the right keywords.
Building this was a technical adventure. I dove into vector graphs and started learning how to tune AI models to handle edge cases and interpret context in genuinely useful ways. Our slack channel was seriously lacking data, so we took what we could get:
