Builds.
AI systems I've led or built, from enterprise product agents to tools I use every day.
Product agents that caught what teams missed
At ThriveAI, we built agents that watched product analytics and support queues, investigated changes, and briefed product teams in Slack.
I worked directly with customers from their first deployment through ongoing use. Product teams already had dashboards and support queues. What they lacked was the time to connect changes in user behavior with what people were saying, investigate the cause, and decide what deserved attention.
A proactive agent has a strange quality problem. A correct observation is useless if the product owner already knows it. Silence is worse when the agent misses the issue that matters. We evaluated findings on whether they were important and new to the person receiving them, then tracked false alarms and misses separately.
At peak, the agents analyzed more than 200,000 product signals a day for teams ranging from scaleups to Fortune 100 pilots. They flagged a mobile release regression within a day of rollout, traced a surge in support tickets to a vendor outage before the customer's team found the cause, and caught broken revenue tracking that the dashboards were hiding. On products with millions of monthly users, anomaly detection stayed below a 2% false-positive rate.
Emmanuel, an AI teammate inside a venture firm
I designed and built Emmanuel for Iterative. He has his own Slack account and inbox and helps the firm support more than 380 portfolio founders.
Emmanuel watches founder-request channels, turns each ask into a tracked ticket, prepares daily briefings, and follows open requests through resolution. He can act inside approved internal systems. Anything leaving the firm waits for a person.
When the team corrects Emmanuel, that decision becomes a standing rule for later runs. The evaluation set has to move too. A fixed golden answer aged quickly as the team's judgment changed, so we kept a versioned baseline of currently acceptable behavior and moved it only after a reviewer confirmed the new result was better.
Giving the agent real responsibility meant someone had to maintain its context, review failures, set boundaries, and update the rules when the work changed.
A French tutor that gets me speaking
I built it after years of language apps had improved my recognition without making me comfortable speaking with my French family. I've given myself until the end of 2026 to reach B2.
It lives in Telegram and teaches through short voice and text exchanges. It gets me producing French in most turns, remembers where I struggle, brings weak material back over time, and maintains a revision page I can use from my phone.