A headless Python rules engine turns complete card-game play into fast, reproducible self-play.
Software engineer · AI systems, simulation & backend infrastructure
I build the systems that let AI agents train, compete, and prove they improved.
I work across deterministic simulation, reinforcement learning, evaluation, and resilient backend automation. Decktrace turns a complex card game into a self-improving capability lab with measurable, inspectable results.
Featured project
Decktrace
An automated capability lab for producing a stronger One Piece TCG policy through self-play, controlled competition, and evidence-bound promotion.
Simulation · Reinforcement learning · Parallel systems · Evaluation
From self-play to a stronger champion.
Five explicit training treatments run against the same fixed game setup. A restartable operator trains them in parallel, cuts weak candidates at controlled rungs, and promotes only after fresh paired evidence clears uncertainty and archive checks.
- Built
- Deterministic headless simulator, shared-action masked PPO, parallel resumable operator, paired evaluation, and automatic champion promotion
- Capability race
- Five arms reach 256,000 steps, three continue to 1,024,000, and two finalists train across a 2,048,000-step horizon
- Measured result
- The first challenger posted a 97.66% point estimate and 96.29% adjusted lower bound, passed archive regression, and made zero invalid actions
- Honest boundary
- The result is confirmed inside the fixed Imu mirror matchup in the simulator—not across arbitrary decks, leaders, or native qualification
One generation · one controlled race
- 01TrainFive explicit arms
fresh + champion lineages - 02ScreenThree survivors
common paired openings - 03FinalizeTwo candidates
2,048,000-step horizon - 04ConfirmArchive-safe champion
automatic promotion
PROMOTED96.29%adjusted lower bound
Five explicit PPO treatments race through controlled rungs while weak candidates stop early.
The first challenger cleared 512 paired openings, an archive check, and zero invalid actions.
Agentic engineering in practice
Automation accelerated the campaign. Explicit boundaries kept the result honest.
Codex and ChatGPT accelerated repository analysis, implementation, testing, and parallel investigation. The training operator now restores its own state, runs independent arms concurrently, checkpoints safely, and publishes a recruiter-readable status view.
I remained responsible for the architecture, experimental boundaries, generated changes, evidence review, and the choice to make game outcomes—not human labels or manual judgment—the authority for promotion.
See the evidence behind that claim