Spotter
An AI advisor, in development, that will talk a college basketball player and their parents through a transfer or player-development decision.
As I put it in September, agents usually make people "put in asymmetrically more work" for what they get back, and these players are "not 24/7 nerds. They play sports." So Spotter asks a player only for what it needs to answer.
Players and parents each get their own version, with its own tone and its own tools. Before I add a tool, I ask myself why it's on my team.
Anything it pulls from the open web comes with a label: "Found on the web. Not reviewed by DepthChartIQ."
Spotter only talks to models through the model gateway I set up and run for DepthChartIQ. If you point it at any other address, it throws an error and doesn't make the call. Its own cost estimate is set to zero on purpose, so the only cost it stores comes from the gateway's ledger entry for that exact call.
Lifting restrictions
Spotter started out locked down. It isn't released yet, and I'm still lifting restrictions one at a time, each one based on real sessions with DepthChartIQ's CEO.
He uses a practice version of the advisor that runs inside Claude, and he plays the athlete himself. Then he writes up what went wrong, edits the advisor's instructions in a fresh Claude session, and keeps or tosses each change on its own. When I reviewed that loop in August I didn't change anything, and the habit I singled out was his keep-or-toss call on every change.
Stack
- Languages
- TypeScript
- Models and agents
- LiteLLMPi agent SDK
- Infrastructure
- Fly.io
See also
Text mePaul Mikulskis
(617) 599-0775If this sounds like your situation, tell me what's going on.
This number usually answers in a minute or two, in my voice, and every conversation comes straight to me.
Text (617) 599-0775