Applied ML portfolio focused on graph ML, predictive maintenance, and practical Python systems.

I build AI-assisted technical projects end to end: experiment setup, training, evaluation, result tracking, and reproducible repo structure.

My strongest public work is around:

• graph ML on OGB benchmarks,
• predictive maintenance and RUL,
• practical Python/Linux tooling when there is a real problem to solve.

I use coding agents as implementation accelerators, not as autopilot.

In these projects my role is to:

• define the problem and target metric,
• choose the direction of experiments,
• run training and evaluation locally,
• compare runs, inspect failures, and decide next iterations,
• keep repos reproducible and readable enough to revisit later,
• publish only work that I can explain and defend.

Typical workflow: idea -> baseline -> run tracking -> error analysis -> next iteration -> final summary

This means the code is often AI-assisted, but the project direction, experiment decisions, triage, and final result ownership are mine.

Repository: MolHIV

Reproducible graph classification project on ogbg-molhiv using PyTorch Geometric.

Highlights:

• official metric: ROC-AUC,
• compared GraphMLP, GIN/GINE, and GINv2 + Virtual Node variants,
• tracked seed variability and checkpoint selection by validation ROC-AUC,
• final N=10 seed batch: 0.762974 +- 0.022686 test ROC-AUC at best-valid,
• ensemble of the same checkpoints: 0.795141 test ROC-AUC.

This repo is a good example of how I work under local hardware constraints and still build a structured experiment pipeline.

Repository: RUL

Remaining Useful Life forecasting on NASA C-MAPSS.

Highlights:

• macro RMSE improved from 46.2155 to 17.0443,
• work focused on leakage fixes, causal features, domain shift, and specialist blending,
• current best system uses a global branch plus FD-specific specialists.

Status:

• strong engineering progress demonstrated,
• project still open, with FD004 remaining the main bottleneck.

Repository: metro3

Recruitment-style predictive maintenance project for early fault warning.

Highlights:

• event recall: 4/4 (100%),
• false alerts/day reduced to 1.925,
• median lead time: 588.9 min,
• includes CLI pipeline, tests, MLflow tracking, and explicit alert policy tuning.

Repository: acer-nitro16-linuwu-fan-curve

Small but practical Linux project: custom fan-curve controller for Acer Nitro 16 on Linux using the Linuwu-Sense driver.

This repo is here on purpose. It shows that I also work comfortably outside pure ML when the task is concrete and system-level.

• practical, experiment-driven ML work rather than tutorial projects,
• AI-assisted implementation with human ownership of scope and decisions,
• repositories with metrics, run summaries, and reproducible commands,
• a portfolio that is still growing, but already shows real iteration and debugging work.

• Python
• PyTorch, PyTorch Geometric, scikit-learn, XGBoost, LightGBM
• RAPIDS (cuDF, cuML), pandas, polars, FAISS, implicit
• MLflow, pytest, Git/GitHub, CLI-first project structure

• LinkedIn: jaroslaw-kata
• GitHub: @benoirczar-ml