Built an AI research desk for Kalshi that can't place a trade unless I click approve (demo money, 3 days in)
I have been messing with Kalshi's demo exchange the last few days and ended up building a local dashboard for it. runs on my laptop and just talks to the Kalshi API.
It pulls every open market and puts stuff next to it that Kalshi doesn't show, like sportsbook lines, live scores, weather forecasts. then Claude (or Gemini, I can switch) researches a market and proposes a trade with its reasoning. I have to approve every one. there are small hard caps on order size and daily spend, plus a kill switch the AI can't touch.
The part I care most about is that I assume the market price is mostly right. every AI estimate gets blended back toward the market price before anything happens, and the model only gets more weight once its raw guesses actually beat the market over enough settled calls. it hasn't earned anything yet. a second model also gives a quick independent number, and if the two disagree by a large margin, it sizes the bet down.
The parts that broke(and a lot did):
- ESPN was rejecting every request for a day, so my "live scores and lines" never showed anything and I didn't notice until I checked the logs
- Kalshi keeps your balance per shard (baseball is on a different one than most things), so my first order failed with "insufficient shard balance"
- my P&L was ignoring fees, for now
The results I have so far: 2 settled, both wins, which means nothing. the demo books are super thin too. in my paper sim, a strategy that buys when sportsbooks disagree with Kalshi only filled a small fraction of what it wanted, so I don't trust demo results to say much about live.
The ai agent is not going live until it can repeatedly beat the market price in at least one category. If you have built something like this, how do you deal with thin books?
And for anyone running something similar: how much edge do you need before a trade is worth it after fees and spread?
Another question I have is what outside data has been actually useful for you? I'm pulling sportsbook lines and weather forecasts, and wondering what else to add"