Over the 2025 holiday break, Jordan built a full-stack trading bot for Kalshi, the prediction market. It scans every active market for whale bets, coordinated entries and volume spikes, then scores each trade. He showed it in Episode 26 (January 9, 2026), along with a real-money result: 8 wins out of 12 trades, and still about $7 down. He did not give an exact build time, only that he built it in his free time over the break.
Prediction markets were one of the big crazes of 2025. Kalshi and Polymarket grew fast, and brokerages and sportsbooks started talking about launching their own. Jordan had seen traders post huge wins, some of which looked like inside information, and wanted to see whether a bot could find an edge without it.
He also admits he is "not really a good gambler" and knows little about sports. He saw that as a plus: the bot would have to work from market behavior, not from his opinions about who would win.
And of those 12 trades, I actually won eight of the 12. So I was actually at about 66% success rate in those trades.
— Jordan Metzner, Episode 26
He still lost money. At 11:34 he explains why:
And I think I lost money for two major reasons. One, the transaction fees are incredibly high, and when I'm taking low margin trades, of like low risk settlements, I found that the transaction fees kinda take away my margin.
— Jordan Metzner, Episode 26
The second reason is that it is "bots trading on bots." If everyone trades on what the crowd is doing, the signal gets priced in. Later in the episode, Jordan asked NotebookLM whether anyone wins consistently with a Kalshi bot. Its answer was arbitrage or providing liquidity, not betting on outcomes. Jordan thinks fees are too high for cross-platform arbitrage to pay off, and expects time arbitrage (trading before the scoreboard updates) to get patched. His verdict: the markets "sure feel like the house, and it sure feels like the house always wins." He planned to withdraw his money and take the $7 loss.
Sam's view: the hard part is finding signals that actually drive earnings, but the build shows what you can do with AI even when the trades don't go your way, and someone with real betting experience could take the same tool further. If you are deciding which coding agent to build something like this with, see Claude Code vs OpenAI Codex.
Jordan connected a full-stack app to Kalshi's API, pulled all active markets into a live feed, flagged whale bets, coordinated entries and volume spikes, and scored each trade with a confidence score. He built it with Claude Opus 4.5, Cursor and Claude Code.
Ours did not. It won 8 of 12 trades, about 66%, but ended about $7 down because transaction fees ate the margin on low-risk trades. Jordan's conclusion was that the only real edge is inside information.
Polymarket was not available to US traders, and Kalshi has a full API that the bot could connect to directly.
Short-dated ones. Jordan focused on live sports because they settle quickly, which gives near-real-time feedback on the algorithm, and because sports have clear winners and losers.
