Muhammad Afaq Jamshaid

Contributed to pandapipes, an open-source Fraunhofer project for simulating energy pipelines, by fixing core bugs, developing a Mapbox-based plotting module, and working on a transient solver for time-dependent flow. Wrote technical documentation and tutorials to improve onboarding and usage for new developers and users of the library. Working on agent-based LLM workflows for domain-specific code generation and repair, integrating sandbox execution, error classification, and adaptive feedback strategies.

Reverse engineered LOGEsoft’s fuel composition software to replace manual range-based search with ML-driven prediction of valid fuel blends for target property values. Designed a hybrid AI pipeline with supervised neural networks, reinforcement learning, and DoE-driven data generation, achieving up to 95% prediction accuracy. Built a surrogate model with iterative learning, conditional retraining, early stopping, dynamic hyperparameter tuning, and real-time feedback on predictions.

- Created supervised fine-tuning datasets of Python, C++, and JavaScript scripts with 1000+ structured prompts and golden responses for SFT and RLHF training. - Benchmarked and evaluated LLM models including ChatGPT, Claude, and LLaMA against scenario based unittest and pytest cases, assessing compliance with the requirements