Erdi Kara - Projects

Overview

I started out as an applied mathematician working on scientific computing and machine learning, often in medical and biological contexts. My academic work included scientific machine learning, PDE modeling, computer vision, statistical data analysis, supported by grants from institutions like the DOE, Simons Foundation and Microsoft ($785K+).

These days at HueDx, I focus more on applied, industry-facing problems, especially computer vision and medical data analysis and diagnostics. The selected projects below include both: academic work that shaped how I think, and practical systems built for real-world use.

HueTools for Colometric Assay Analysis

HueTools is a software platform I co-developed for the quantitative analysis of paper-based colorimetric assays using smartphone imaging. It integrates a mobile app and web backend to standardize image capture, apply precise color correction, and support interpretable machine learning models for concentration prediction. HueTools supports LoB, LoD, and LoQ estimation and has been validated on both lateral-flow (LH) and vertical-flow (ALT) assays. It's already deployed internally at HueDx, powers the newly launched Hue Fertility platform and will be made publicly available soon. A detailed technical paper is currently under review.

Multi-Object Tracking

Led the development of a high-performance object tracking library, achieving 98% tracking accuracy outperforming state-of-the-art models (StrongSORT, Deep OC-SORT) for the tested experiment. Turned out to be you can actually use it for variety of different tasks as well. Check these links. Paper Link | GitHub.

Optimizing CAR T-cell Therapy

Led the development of a mathematical model integrated with statistical data analysis to study how we can improve the effectiveness of engineered T-cells via nearby bystander cells. Check out here to see some interesting results. Paper Link | GitHub.

Applied LLM and Vision LM Projects

These projects focus on document understanding, retrieval, and model alignment of LLMs.

Interpretable RAG for Complex Documents : a multimodal RAG system (ColPali + Qwen2-VL) for querying images, tables, and text in PDFs. Github
PDF Search & Q&A System : a natural language interface for querying large PDFs using LangChain, Sentence-T5, FAISS, and LLaMA-2. Github

Fine-Tuning & DPO Alignment for LLaMA-3.2: a full training and alignment pipeline using Unsloth and DPO; improved performance by 60%. Github

Coupling FEM and PI-DeepOnet

This is something I’ve been working on for a while, trying to bring together the best of two worlds: classical scientific computing and operator learning. The idea is to couple finite element methods (FEM) with physics-informed DeepONets to build data-efficient surrogates for PDE-based systems Paper Link | GitHub.

Publications

HueTools: A Modular Image Processing and ML Toolkit for Quantitative Colorimetric Assay Development, Nidhi Mennon, Erdi Kara, et al. 2025, Under Review.
Physics-Informed DeepONet Coupled with FEM for Convective Transport in Porous Media with Sharp Gaussian Sources, Erdi Kara, Panos Stinis. 2025. Arxiv preprint.
Mathematical Modeling Insights into Improving CAR T Cell Therapy for Solid Tumors with Bystander Effects, Erdi Kara, Trachette L. Jackson,et al. 2024. Nature Systems Biology and Applications.
Deep Learning Based Object Tracking in Walking Droplets and Granular Flow Experiments, Erdi Kara, J. Nathan Kutz, Aminur Rahman, et al. 2023. Journal of Real-Time Image Processing.
Tumor Ablation due to Inhomogeneous–Anisotropic Diffusion in Generic 3-Dimensional Topologies, Erdi Kara, Aminur Rahman, Eugenio Aulisa, 2021. Physical Review E.