Detection & Counting
Finding and counting objects in crowded, messy images — microscopy cells, scientific captures, anything where manual counting breaks down.
Etele Kovács — Computer Vision & ML
I build computer-vision and ML systems that take raw data all the way to working predictions.
I work in computer vision and applied ML. I've built microscopy tools that researchers use daily — counting pico-algae cells and measuring bacterial biomass straight from the microscope — taking each project from raw data to an installed, working app.
Projects
Four projects, documented end to end.
Real data, real metrics, and the tradeoffs written down — including the parts that didn't work. Where the code is public, it's linked.
Microscopy Image Analysis
Pico-Algae Detection and Counting
A microscopy detection pipeline and offline desktop app that count pico-algae cells from paired brightfield and fluorescence image channels — in daily use at the lab bench.
Result Snapshot
Processed Samples
250
Applied MLComputer VisionObject DetectionMicroscopy
Case studyOpen
Microscopy Image Analysis
DAPI Microscopy Biomass Pipeline
From DAPI fluorescence microscopy to biomass: a fine-tuned Cellpose model, a faithful port of the Zeder biovolume algorithm, and an offline desktop app colleagues use to measure biovolume straight from the microscope.
Result Snapshot
Status
In use
Computer VisionImage SegmentationMicroscopyCellpose
Case studyOpen
Natural Language Processing
NER on Code-Switched Spanish-English Tweets
Named entity recognition on tweets that mix Spanish and English mid-sentence — with a lightweight language-aware bias layer that beat plain fine-tuning of mBERT and XLM-RoBERTa.
Result Snapshot
Best Span F1
0.649
Group ProjectNLPNamed Entity RecognitionCode-Switching
Case studyOpen
Risk & Tabular ML
Automobile Insurance Claims Risk Modeling
Predicting motor-insurance claim frequency on 678K French policies — with a decision tree, neural network, and PCA implemented from scratch and benchmarked against a Negative Binomial GLM.
Result Snapshot
Policies Analyzed
678,013
Group ProjectMachine LearningFrom ScratchTabular Data
Case studyOpen
Work Explorer
Compare the projects at a glance.
Switch between projects to see the problem, the key numbers, and a link straight into the full case study.
Microscopy Image Analysis
Pico-Algae Detection and Counting
An end-to-end deep learning workflow for detecting and counting pico-algae in microscopy imagery, now shipped as an offline desktop app that colleagues run at the microscope. The core model is a custom 6-channel Faster R-CNN that fuses paired microscope images; around it sit training, tuning, and post-processing sweeps, plus a local review UI where corrected detections are exported straight back into training data.
Processed Samples
250
Self-annotated in CVAT before training.
Labeled Boxes
16,181
Applied MLComputer VisionObject DetectionMicroscopyScientific ImagingPyTorch
Open case studyAreas
What I work on.
The kinds of problems the projects above cover — from detection and counting to full data-to-deployment pipelines.
Classification
Turning structured or visual data into reliable predictions, with evaluation that holds up outside the training set.
Segmentation & Analysis
Image-analysis pipelines built to be inspected: clear outputs, reproducible runs, and outputs that fit the scientific domain.
End-to-End ML Systems
The whole path — preprocessing, feature engineering, APIs, metrics, and a project structure that's ready to deploy and hand off.
Skills & Stack
The stack behind the work.
Every tool here is one I've shipped real work with — not a wishlist.
Vision & Modeling
PyTorchTorchvisionOpenCVHugging Face Transformersscikit-learnCatBoost
Data & Scientific Workflows
PythonPandasNumPySciPyMatplotlibpvlib
Interfaces & Delivery
FastAPIPydanticReactViteTypeScriptVercel
About
How I think about the work.
Serious engineering, documented plainly — so you can see exactly what was built and how it holds up.
I take computer-vision and applied-ML problems from raw data to something that runs — and I document exactly how each project works, what it measures, and where the tradeoffs are.
My approach
Reproducible pipelines and clean interfaces that make the work easy to inspect, trust, and build on.
Read moreContact
Get in touch.
Want to know more about any of this work?
I'm happy to talk through any of the projects here — the decisions behind them, the parts that didn't work, or the details that didn't fit the write-ups. Email and LinkedIn are the best ways to reach me.
View CV