RLTriage

RLTriage is a web app for reinforcement learning teams who keep losing weeks to “it doesn’t learn” debugging. You upload training logs (TensorBoard, Weights & Biases exports, or simple CSVs) plus a minimal config snapshot, and the app runs a battery of automated diagnostics: reward scale/pathologies, action saturation, observation drift, exploration collapse, replay buffer issues, and environment non-stationarity signals. It then produces a ranked list of likely root causes with concrete fixes (e.g., normalize rewards, clip advantages, change entropy schedule, fix terminal handling) and a reproducible “next run” checklist. It’s not magic: it won’t invent a better algorithm for your problem, but it will reliably catch common failure modes and reduce guesswork. The product is positioned as an engineering tool, not a research platform—focused on saving iteration time and compute spend.