Etele Kovács

Research project · Python · Cellpose · OpenCV · NumPy

• Building a proposal-to-biomass pipeline for DAPI fluorescence microscopy: high-recall segmentation proposes candidate objects, crops are exported for review, contours are refined, and biovolume is measured only from the clean contours — in the spirit of the Zeder/YABBA phytoplankton tools.
• Implemented and unit-tested the segmentation, proposal generation, crop export, annotation manifests, image-level dataset splitting, and local refinement stages behind a single CLI, with Docker/RunPod GPU training set up.
• Ongoing: the crop classifier and full end-to-end orchestration are not finished yet.

Python · PyTorch · OpenCV · Computer Vision

• Built a machine-vision pipeline with OpenCV and a 6-channel Faster R-CNN (ResNet50-FPN) to detect and count pico-algae in paired microscopy images, where dense frames with overlapping cells break standard approaches.
• Tuned the full post-processing stack (anchor sizes, NMS, confidence thresholds) to cut counting MAE from 3.94 to 2.42 across 16,181 annotated objects in 250 self-annotated images.
• Owned the dataset end to end: designed the annotation format in CVAT, ran preprocessing, built reproducible splits, and validated across imaging conditions.
• Optimized for accurate counts rather than perfect boxes — used count MAE instead of mAP, which gave a far clearer signal during iteration.

Python · scikit-learn · FastAPI · React

• Modeled spectral light-attenuation coefficients across 380-900 nm from ~245 real water samples, using chlorophyll, CDOM, and TSS as inputs with PCA to handle spectral correlation.
• Reached R² 0.86 and RMSE ≈ 0.65 (log scale) on the held-out test set — robust given the noise typical of environmental water data.
• Built the full stack: a FastAPI backend for predictions and a React/Vite frontend for visualizing irradiance at depth and exporting results. Wanted it usable, not just a notebook.
• Kept documentation tight throughout so the pipeline could be handed off and reproduced without digging through code.

Group project (ITU) · Python · NumPy · scikit-learn · PyTorch · statsmodels

• Three-person BSc exam project predicting motor-insurance claim frequency on ~678,000 French policies. I built the feed-forward neural network from scratch in NumPy (custom losses, optimizers, batching), the PyTorch reference MLP, PCA, and the data preprocessing/EDA pipeline.
• Across the team's models, a Negative Binomial GLM won on test RMSE (3.03) and Poisson deviance (1.88), beating the trees and neural net; a Random Forest reached ROC AUC 0.65 on the binary claim task.
• Takeaway: on noisy, zero-inflated claim data the simple statistical model matched or beat far more complex ones — match model complexity to the signal, not the hype.