Case study
AI Sports Research Engine
I built an agent that could research sports betting markets on its own, and it found a consistent edge. I killed it because the edge did not grow but the cost of maintaining it did. The machine I built here became ProfitLabs.
The problem
Sports betting markets are not perfectly efficient. They are close, but there are consistent places where eleven seasons of record disagree with what the market prices in, by a percent or three. The question was whether a model with the right tools could find those places faster than I could, and whether the edge was worth the cost of running the system.
What it did
Claude with 22 SQL tools over 11 seasons of NBA and NFL data. Instead of me writing every query, the model could ask its own questions, read what came back, and follow up. My job was writing good tools and deciding what to do with the findings. It surfaced systematic biases: the gaps between what the market priced and what the record showed, holding across enough seasons that it was not noise.
The decision to kill it
The edge was real and it was consistent. The problem was everything around it: nightly ingestion, dashboards, and models maintained separately for every sport. Each league I added raised the running cost. None of them raised the edge.
Prediction markets were a better version of the same bet. Mispricings on Polymarket and Kalshi were larger. Liquidity was deeper on events nobody had a clean model for. And the data underneath, whale flow and the same contract priced two ways on two venues, was something I could own rather than rent. The technical core carried over unchanged: an agent reading structured context out of a Postgres warehouse. I kept the machine and pointed it at a market worth pointing it at.
I made the call on January 29, 2026. ProfitLabs went on to track around $120M in daily volume for 5,000+ users.
What the cost actually looked like
Every league needed its own data pipeline: different stat providers, different schemas, different update cadences. NBA games happen every night during the season. NFL is weekly but the prop market is deeper. I was maintaining two separate ingestion jobs, two sets of features, and two model configs. Adding a third league (MLB was the obvious next one) would have meant a third of everything, and the edge would have stayed at 1 to 3%.
The economics were simple: each league cost roughly the same to maintain but did not compound the return. Prediction markets, by contrast, had one data source (the exchanges themselves), larger mispricings, and the same core infrastructure.
Python, Claude API, SQL, Next.js.
No longer running, January 2026.
Back