RSIHub
I drove RSIHub’s development as lead developer and led its 0→1 implementation. My work spans the experiment lifecycle and framework architecture, evaluator isolation and mutation boundaries, candidate lineage and evidence retention, the recipe and operator system, runtime and Harbor integration, benchmark infrastructure, testing, and documentation.
RSIHub enables controlled agent self-improvement while keeping the evaluation mechanism outside the candidate’s mutable surface and preserving reproducible evidence for each generation.
Across four self-improvement strategies and all 186 tasks in two benchmark suites, the best configurations improved Terminal-Bench 2 with MiniSWE from 56.2% to 69.7% (+13.5 points) and Tau3 Banking with Codex from 24.7% to 34.0% (+9.3 points).
