About
I have worked with data at three companies and across sports, manufacturing, finance and biomedical projects. My focus is on understanding what the data contains and what it means for the people who rely on it. I use prompt engineering and AI tools to reach answers faster, and I review every result carefully. I follow GDPR and the FATES framework, so the analysis is both responsible and accurate.
Featured · Recognition
I was nominated by the University of Sheffield Careers & Employability Service for my work at TrackMyMachines Ltd, a project funded by the South Yorkshire Innovation Program.
My team built SwimMetrics, a dashboard that shows how ready athletes are for the Olympics, using more than 2M historical athlete records. We built it in a single 72-hour sprint.
“Tarun is a driven individual with a thorough and considered approach to his work. He delivered a project for TrackMyMachines that required deep understanding of our system, data and methods. Not only did he spend time trying to understand our business needs but his research and work helped us to see alternatives for our product that will ultimately bring more value to our customers.”
Experience
TrackMyMachines Ltd · University of Sheffield
Sheffield, United Kingdom
Department of Computer Science & Engineering, PES University
Bengaluru, India
Third Axis Ltd
Bengaluru, India
Education
The University of Sheffield
Sheffield, United Kingdom
My coursework covers statistical learning, optimisation, large-scale data engineering and applied machine learning. The university is known for its research in data, computing and biomedical sciences.
PES University
Bengaluru, India
Specialised in aerospace.
Tech Stack
Projects
Earlier studies reported high accuracy for predicting breast cancer stage, but did not check whether the SHAP explanations behind those models were stable across different model types.
Predict cancer stage from the RNA-seq data of 1,072 TCGA-BRCA patients, and test whether the explanations agree, instead of assuming a good SHAP result means real biological signal.
I used log2 normalisation, variance filtering and PCA, then a leakage-free pipeline with stratified 60/20/20 splits and five-fold tuning. Six classifiers, including a PyTorch network wrapped as a scikit-learn estimator, were tuned and tested the same way, with 95% bootstrap confidence intervals. SHAP explained each one, and I measured agreement by top gene overlap, Spearman correlation and Kendall's W.
None of the models could reliably separate early from late stage cancer. Their explanations, however, agreed strongly within each model family, which matters for how post hoc explainability is used in clinical AI.
Contact
I am open to data science and machine learning roles, and happy to relocate for the right team. Email is the quickest way to reach me, and I usually reply within a day.
