Tom Muga

Rebuilt a Python stochastic simulation core in C++ to achieve 100× throughput improvement versus the original implementation. Also reduced experimental search space by 90% and demonstrated cross-instance generalisation by learning intervention dynamics from one instance and generalising to 998 additional cases.

Delivered 100× throughput improvements by replacing Python computation cores with purpose-built C++ implementations, validating gains with benchmarks and hardware counter analysis. Built performance-critical simulation and optimization systems (C++/Python interoperability, parallel execution, and numerically stable algorithms) and shipped verified PRs with reproducible results.

Improved a Monte Carlo simulation engine to reach 2.38× throughput over a Python baseline by restructuring parallelisation and eliminating intermediate storage. Fixed a parallel pragma placement bug, used a single-pass Welford algorithm for numerically stable mean/stddev, and added risk estimation (expected revenue and probability of negative cashflow).

Built a fixed-capacity belief distribution framework with a compile-time memory pool to eliminate dynamic allocation and reduce fragmentation. Implemented a certainty-aware two-objective Dijkstra (biasing toward higher certainty) and reduced Section size from 32 to 24 bytes by member reordering and replacing std::string with char.

Rebuilt a Python stochastic causal simulation core in C++ to achieve 100× throughput improvement versus the original. Implemented a contiguous, allocation-minimal C++ engine with a Python orchestration layer for configuration and aggregation.

Optimized an entropy-guided stochastic optimizer to reach 22.38× throughput improvement through profiling-guided hot-path changes. Reduced cache miss rate from 53.64% to 12.15% (−89% absolute misses), with instruction and branch misprediction reductions, while improving numerical stability with regularized Cholesky.

Built a zero-serialization, zero-copy WebAssembly execution model using C++ as the memory owner/scheduler, C# as a stateless compute module, and minimal JS bootstrapping. Achieved stable 60fps at 1,024 entities with 99.95% frame consistency via shared WASM linear memory and a single JS boundary crossing per frame.

Implemented a multi-stage optimization pipeline (Gaussian sampling, evaluation, hierarchical clustering, adaptive refocusing) with numerically stable Cholesky decomposition using jitter. Improved throughput by 22.38× and reduced cache miss rate from 53.64% to 12.15% (−89% absolute misses) via targeted hot-path changes.

Modified IS-MCTS selection to incorporate action-level variance into the UCB decision rule for risk-adjusted search. Empirically showed variance penalisation amplifies first-mover advantage with statistically significant divergence from the canonical implementation.

Identified and refactored a redundant two-pass hot-path loop (rcCalcBounds) into a single-pass fused min/max operation. Improved throughput by ~10% (106 ns → 95 ns) and reduced instruction and cycle counts per hardware counters.

Improved a Monte Carlo simulation engine by 2.38× over a Python baseline and an additional 1.71× by removing intermediate storage and restructuring parallelization. Fixed a structural OpenMP bug that made the MC loop sequential and used a fused single-pass Welford algorithm for mean/standard deviation.

Completed a Software Engineering Diploma in 2022.