Experiments started out of curiosity
These were small experiments to try model-driven decisions in situations where speed matters. I explored two settings: football recommendations and a coin-trading bot. Both stayed at the proof-of-concept stage.
Football: a recommendation can become stale
The football prototype received the ball’s position and player coordinates. It produced recommendation scores for where the eleven players on one team should be positioned and where the player with the ball should pass.
I researched tactics and papers to build a decision harness: a set of criteria for the model’s scores. The design also considered human reaction time, model inference, and delivery latency. A recommendation was intended to be delivered only when its validity margin remained sufficient after those delays.
Finding matching video and coordinate data was difficult. I assembled a limited test using a short clip and player coordinates found online. I do not have football expertise, and I have not established whether the recommendations were tactically useful or whether the filter improved their quality.
Trading: decisions inside a simulation
The trading prototype asked a model to choose buy, sell, or hold using another research-informed harness. A rough backtest explored what might have happened if those decisions had been followed. The result was not satisfactory to me, and I did not proceed to real trading.
The prototype included an order-integration path, but no live trades were made. Whether an order would actually fill remained a difficult part of the simulation. I limited the experiment to larger coins under a small-order market-impact assumption; that assumption was not independently validated.
What the experiments explored
The work raised questions about how long a recommendation stays relevant and how much a simulation depends on its assumptions. These are limited prototypes, with no established recommendation quality or profitable trading result.