Architecture
Distributed trainer/worker topology, checkpoint sync, stale rollout rejection, failure handling, scaling points, and hands-off self-healing cloud operation.
Read architecture docCore Docs
Headless Simulator & Search-Combat
The June pivot to the sts_lightspeed C++ simulator, the variant series that proved combat execution was the universal wall, and the search-combat + diagnosed-macro stack that produced the first winning agents (mainline_v3).
Headless Cloud Deployment
Running the worker/trainer stack headless on a GPU-less GCP spot VM: one-shot installer, Xvfb + software GL, Java 8, spot preemption recovery, and the nine bring-up challenges with fixes.
Read deployment notesTechnical Writeup
Implementation detail on observations, actions, reward shaping, behavior cloning, PPO, parallel training, and limitations.
Read technical writeupExperiment Reports
The full arc, from BC baseline and live-game PPO through the simulator variant series to the first winning agents (009–011), backed by the public experiment registry.
Open experimentsScripts Reference
Plain-English guide to each training, evaluation, environment, plotting, and logging script.
Read script guideDashboard
Static results viewer with embedded public metrics and support for loading local CSV files.
Open dashboardPortfolio and Resume Framing
Recruiter-facing summary, public demo assets, and resume-ready bullets.
Open portfolio page