✦ Building a production-style C++/Python Monte Carlo engine for American-style derivatives pricing — Longstaff–Schwartz initialization, deep-RL exercise policies, and scored validation against published academic benchmarks (arXiv:2606.17545)

✦ Latency & Throughput Optimization: Engineered a cache-aligned C++20 GBM simulator (Structure-of-Arrays, OpenMP) with vectorized PyTorch inference and RL short-circuit logic, cutting combined basket-put benchmark runtime from ~2.8 hours to ~40 seconds (120–160×) while improving price accuracy toward Table 3 targets

✦ Pricing Accuracy & Reliability: Architected a two-stage optimal-stopping pipeline (LSMC → ADNN → adaptive grid RL) with validation-guided epoch selection, target-aware checkpointing, and deterministic forward-MC scoring on fixed path banks — preventing RL regression from degrading in-band Stage 1 models

✦ Validated Benchmark Suite: Reproduced 3/7 published CARLOS contracts within tolerance (B1, B2, M2.A PASS); M2.A priced within 0.002 of target (14.171). Designed CI-ready pass/fail benchmark gate with documented scoring protocol (ADRs) and pytest regression suite