- Implemented batch prompting and attention masking using PyTorch for local model batch evaluation, increasing throughput by over 400%- Programmed various experiments and evaluation pipelines to compare various LLM tuning strategies (DPO, RLHF, and Contrastive-SFT), providing insight into what are the best practices for improving alignment and safeguarding against prompt injections- Designed/Coded 10+ experiments using Python HuggingFace to analyze TinyStories-33M LLM model's performance after training on various supervised learning(SL)/RL splits; Concluded that SL was roughly 15% more effective than RL on simple language tasks

CoreGraphics/ML

- Used Pyserial/Shell to program robust automated board testing software, increasing circuit board testing efficiency by 1.6x- Identified/fixed bugs in deprecated Python product testing software, speeding up medical device production by 1.8x- Engineered C++ code for real-time medical procedure data logging; integrated code into large existing C++ QT GUI application, vastly streamlining system error tracing- Prototyped hospital image database interface feature on an AWS EC2 server; presented findings to my 3 team members

- Programmed board updater software for 3 different microcontrollers (MC) using PySerial and shell scripts (to execute Linux CLI commands), successfully updating 300+ devices in the field- Designed/conducted software/physical unit tests, ensuring updater functionality prior to release- Implemented a MC bootloader using C and integrated commands into the pre-existing system, making field updates possible