Dhilan Avula

University of Texas at Austin

Aug. 2026 – Present

Most reinforcement learning needs huge amounts of trial and error, which is a problem when every trial is an expensive physics simulation. The lab built dfPO (NeurIPS 2025), which reframes RL as a continuous-time control problem so a physics prior is built into the learning itself. I'm reproducing it from the paper's code and benchmarking it against 12 standard RL algorithms, like PPO, SAC and TRPO.

• Getting dfPO and 12 Stable-Baselines3 baselines running in Python
• Running benchmarks on 3 scientific tasks: surface modeling, grid-based PDE control and molecular dynamics
• Learning the optimal control and Hamiltonian mechanics math the method is built on

University of Pennsylvania & Adelaide University

2026 – Present

Coding every minute of a classroom recording for collaborative problem solving by hand doesn't scale, so the team uses open-weight LLMs to score each minute against 19 sub-skills. The question I proposed, and the team adopted: when the model also sees earlier minutes of the transcript, does it label more accurately, or does it just label more? I'm comparing the two setups and checking whether the extra detections are backed by evidence the model actually had.

• Comparing 6 open-weight LLMs across two setups: current minute only, or current minute plus earlier context
• Analyzing 16,800+ model outputs in Python and pandas, merging two different export formats into one dataset
• Collaborating across U.S. and Australian time zones through a shared GitHub repo

University of North Texas · Biomedical Engineering

Organoids are small 3D clusters of cells grown in the lab that behave like a simplified version of a real organ, so researchers can study disease and drug effects without using patients. I help prepare organoids for the lab's cardiovascular experiments.

• Help prepare organoids for experiments alongside the lab's graduate researchers
• Learning wet-lab technique and how tissue models connect to the cardiovascular modeling I do on the computational side

UT Arlington College of Business & UF Warrington College of Business

May 2026 – Present

Banking research often has to combine datasets from several sources, and the same bank or company can show up under different names, spellings and IDs in each one. Entity resolution is the work of figuring out which records are actually the same entity. I do that matching across large multi-source financial datasets for a finance professor's empirical banking study, then verify the matches so the classifications the paper relies on are accurate.

• Entity resolution across large, multi-source financial datasets
• Verified matches to improve classification accuracy before analysis
• Prepared data with a graduate collaborator for a paper and a fall presentation at UT Arlington

ATLAS Research Scholars

Oct. 2025 – Present

Heart disease risk calculators mostly rely on standard clinical numbers like blood pressure and cholesterol. I was selected for ATLAS through an interview and developed my own question: do factors outside that list, like psychosocial stress, socioeconomic status and inflammation biomarkers, make cardiovascular disease prediction models more accurate? I'm now writing it up as a paper.

• Developed the research question and compared published cardiovascular risk prediction models
• Presented at MAA MathFest and Harvard Medical School's Women's Health Conference
• Preparing a paper for undergraduate and high school research conferences

Women's Health Conference, Harvard Medical School

Oct. 1, 2026

Most heart disease risk models are built on clinical numbers like blood pressure and cholesterol, and they often miss the socioeconomic and psychosocial factors that shape women's risk. With four other ATLAS scholars (equal contributors), I searched PubMed, Scopus and Embase, screened 2,770 records down to 28 studies, and compared how well current models predict risk for women.

• Systematic review: 2,770 records screened, 28 studies included
• Compared discrimination and calibration of conventional scores against models adding biomarker, socioeconomic and psychosocial data

MAA MathFest, Boston

A poster from my ATLAS research at MathFest's undergraduate research conference: a quantitative comparison of about 40 cardiovascular disease risk prediction studies.