Same standards. No one else’s roadmap.
My own company — where I build AI-native software end to end: product, architecture, infrastructure, billing, and the unglamorous parts in between.
An AI teammate that runs a team’s time off and shifts from inside Slack.
Live · hundreds of users
Teams ask in plain language — book next week off, swap a shift, who’s out Friday — and Officebot handles it against live calendars, with Google Calendar sync, shift planning, and daily summaries. Used by teams at companies including Cisco Meraki.
getofficebot.comHow it ships
01
Evaluation before features
Every PR runs offline prompt and model accuracy tests in promptfoo. Nothing reaches customers without passing them.
02
Judge production, not the demo
Langfuse traces live traffic, with LLM-as-judge scoring of tool-calling accuracy against real data.
Three eras, one product
- era 1
Regex
It started as pattern-matching commands. It worked — as long as people typed exactly what it expected.
- era 2
NLP
Intent and entity models let people speak naturally. Better, but every new request type meant new training data.
- era 3
Agentic
Today an agent on the Vercel AI SDK and AWS Bedrock plans the steps and calls the tools itself — one sentence can book a whole week.
An AI study companion that adapts to the learner.
In progress
The idea came from my daughter, a high school student, and we’re building it together. Under the hood: LLM-based personalization and adaptive learning workflows.