Samuel’s Personal Site

Role: Software Engineer Intern

I wrote tests to ensure that the company’s robot operates as intended, which included moving in the correct direction, and being able to recognize and parse data from its peripherals. Doing so granted us assurance individual components’ behaviours (or at least the ability to trace when something goes wrong). Overall it was a good experience working with other high-performing individuals, to build things that directly interact with the physical world.

Role: ML Engineering & Research Intern

For ML Research, I had control over the entire training pipeline (preprocessing data, feature engineering and model design, model training). I explored using existing pretrained models as embedding sources for feature extraction, and read a lot of papers on this. On top of that, I explored the feature distribution within our datasets to look for potential imbalance or under-representation of some scenarios.

For ML Engineering, I wrote and deployed code to a C++ video analytics pipeline. I designed data structures and algorithms that fuse information from different inference services that run on varying frame-based and time-based periods. I had extensive discussions with others too to ensure that various corner cases would be handled appropriately, to deliver accurate safety insights while not causing alert fatigue from spam. I also implemented an async wrapper around Nvidia’s Optical Flow Library which required wrangling with recycled buffers (NVMM memory management), the CUDA API, and thread-safe state management. Additionally, I performed investigations into discrepancies in performance between our research notebooks and in production, which was a trial by fire towards familiarizing myself against the Nvidia Metropolis framework.

Role: Backend Engineer & Data Analyst

SBL was engineered from the ground up in 2024, with new gamified components being incrementally added since. In August 2026, we had 2.2k+ high-school, local students participating, of which a large proportion do not take Biology and hence use our League as an opportunity to gain exposure. Building a large-scale, synchronous contest is not easy - we get over 10 million hits to our DB in the span of just four hours - but through extensive discussions and testing, we have had a clean track record with no unexpected behaviour mid-contest.

I also perform data analysis on participants’ submissions, which includes breaking down performance across questions (via Rasch fits) to estimate their implied difficulty, finding suspicious event chains indicative of collusion/cheating, and finding miscellaneous fun and useful stats for both participants and the Organizing Team. Stay tuned for more stats coming soon!