Drekken walks through his telemetry dashboard — a tool he built to stop guessing why his bot lost and start knowing. Using STDERR/STDOUT logs piped into DuckDB and visualized through Streamlit, he layers bot decision data onto replay analysis (via SC2 Replay Parser) to pinpoint exactly where things went wrong: economy stalls, army value divergences, scout intelligence, and decision confidence over time.
Key Takeaways:
Telemetry replaces guesswork — When your bot loses, print statements tell you what happened, telemetry tells you why. Structured event logging into a queryable database beats scrolling through log files every time
Layer bot decisions on replay data — Combining what your bot was thinking (confidence, strategy switches) with what actually happened (army value, worker counts, resources) reveals the exact moment things went wrong
Battle sim confidence is a leading indicator — Drekken’s bot showed declining confidence around the 3-minute mark, well before the loss was obvious on the score screen. That’s your early warning system
DuckDB + Streamlit is a lightweight stack — No need for heavy infrastructure. Output structured events from your bot, store them in DuckDB, query with Streamlit or Jupyter. You could also use JSONL locally
Pinpoint the exact failure moment — Instead of scrubbing through a 10-minute replay wondering what happened, telemetry tells you “at 4:57 the opponent’s army value skyrocketed” so you know exactly when to look
Scout intelligence integrates into the timeline — What your bot saw (gateway, nexus, stalker, void ray) appears alongside economy and army data, giving you the full picture without loading the replay