field notes
krrish kapoor
Hey, I’m Krrish. I’m a product analyst at Baselane in Toronto. On the side I build AI agents, mostly to see what works, what breaks, and what I can learn from it.
Photo 1 of 5: Krrish sitting on concrete stairs in a dark suit
Outside of work it’s books, tennis, F1, and Bleachers. I also build little systems for my own life, from habit trackers to trading bots. Some of them stick, most of them teach me something, and a few were just an excuse to build.
This is where I write down what I’m building, reading, and figuring out, usually while it’s still a bit broken.
Experiments
Things I’m building
AI agent experiments
I build personal AI agents to see what works, what breaks, and what I can learn.
Slack data assistantTeam hackathon
My team won the internal hackathon with a bot that answers data questions in Slack.
Fraud detection model
I also once trained a fraud model that scored a perfect F1 and it turned out to be cheating. That one taught me more than the stuff that worked.
Earlier experiments
- Customer churn predictionFeature engineering, model comparison, and business analysis for finding which customers may leave and what a retention team could do next.case study
- OAS retirement strategy simulationA Monte Carlo model testing when Canadians should draw OAS under different lifespans, account types, tax rules, and market conditions.case study
Tokens
My AI throughput
Daily tokens across my AI tools and agents.
- Tokens, last 7 days
- 1.45B
- Streak
- 23 days
Since Sep 1
see allCollected writing
Notebook
Notes, ideas & longer writing
- Building a self-hosted multi-agent runtime
Why I built my own agent system around the Claude Agent SDK, and what memory, specialist agents, recovery, and trust changed along the way.
- What ROC-AUC hid in my fraud-detection project
A high score told me the model could rank transactions. It did not tell me how many alerts a fraud team would have to review.
Bookshelf





