Dragoș Boncan

Dragoș
Boncan

I build machine learning systems and real-time 3D simulations.

Computer Science graduate (GPA 3.83/4), now studying for an M.Sc. in Data Science at Politehnica Bucharest. I'm looking for a junior role in machine learning, data science or software engineering, full-time or part-time.

See the eclipse simulator Download CV
GitHub LinkedIn dragos.boncan@gmail.com
A solar eclipse in the 3D Solar Eclipse Simulator

More projects

Each one pairs a concrete problem with a measured result.

CannyCore

Parallel edge detection. Team of 3, Oct 2025 to Jan 2026

Repository on GitHub

We implemented the full Canny edge-detection pipeline in C++ and parallelized it in eight variants: Pthreads, OpenMP, MPI, CUDA and four hybrids of them.

Speedup over the serial version on an 8000 × 8000 image
OpenMP3.0×
Pthreads3.4×
MPI + Pthreads6.4×
CUDA44.6×
  • On the GPU, one thread handles each pixel, the Gaussian blur uses shared memory and hysteresis is rewritten without recursion. An 8K image takes 0.18 s instead of about 8 s.
  • Profiling with Intel VTune, perf and Nsight Compute showed the Gaussian blur took 48.6% of the serial runtime and MPI broadcasts took 36.8% of the pure-MPI runtime. I used this to design the hybrid variants (halo rows, separable blur).
  • Across 7 image sizes, CPU speedups level off because the hysteresis step is sequential (Amdahl's law) and because of communication overhead.

Built with C++, CUDA, OpenMP, MPI, Pthreads, Intel VTune, Nsight Compute, perf, Make, Git.

Visual classification and linguistic analysis

Computer vision and NLP, Jan 2026

Repository on GitHub

Two deep learning problems, built and compared in PyTorch.

  • Compared a ResNet9 CNN with a HOG + MLP baseline on two 10-class image datasets. The CNN reached 97.2% accuracy on satellite images (26 points above the MLP) and 87.0% on ImageBits (17 points above).
  • Fine-tuned a pre-trained model on only 2,000 satellite images and reached 92.4%, using Mixup and CoarseDropout augmentation.
  • Trained LSTM and RNN models for Romanian sentiment analysis on 28,946 reviews with FastText embeddings, reaching 88.5% accuracy (F1 about 0.90). The LSTM trained far more reliably than the RNN.
  • Automated the analysis with 7 Python modules that produce EDA, t-SNE, word clouds and training charts, and documented everything in a 140-page report.

Built with Python, PyTorch, Albumentations, FastText, AdamW, OneCycleLR, scikit-learn, Matplotlib, pandas.

Predictive regression systems

Statistical modeling, Nov 2025

Repository on GitHub

I benchmarked 7 regression models with 5-fold cross-validation on used-car prices (23,734 Autovit listings) and bike-rental demand (10,886 records).

  • For car prices, a polynomial SVR beat the linear baseline (R-squared 0.94 against 0.84). Random Forest overfitted: 0.95 on training data, 0.79 on test data.
  • For bike demand, Gradient Boosting reached an MAE of 22, against 41 for Linear Regression.
  • Adding recent-history features (previous day and week, rolling averages) raised Random Forest to an R-squared of 0.98 and an MAE of 6.8.
  • Data preparation included EDA, missing-value imputation, Isolation Forest outlier removal, log-transforms and 34 engineered features.

Built with Python, scikit-learn, pandas, NumPy, seaborn, joblib.

AI medical diagnosis assistant

Semantic search, Jun to Sep 2025

Repository on GitHub

An educational prototype, built during the 3DPUB Summer School, that turns free-text symptoms into a list of likely conditions. It uses Sentence Transformers embeddings, K-Nearest Neighbors with cosine similarity and PCA for speed. Model caching and an offline TF-IDF fallback keep it working without a network, and a desktop chatbot (Gradio and PyWebView) wraps it.

Built with Python, Sentence Transformers, scikit-learn, pandas, NumPy, joblib, Gradio, PyWebView.

Sokoban solver

Search and optimization, Mar 2025

Repository on GitHub

A push-only Sokoban solver using LRTA* and obstacle-aware A* and BFS heuristics, with deadlock detection. It solves 8 of 9 benchmark maps in under 14 seconds each.

  • LRTA* explored up to 13 times fewer states than Simulated Annealing (5,351 against more than 70,000).
  • Solution paths were 48 to 87% shorter, 69% on average. A Tkinter viewer replays each solution step by step.

Built with Python, Tkinter, A*, BFS, LRTA*, Simulated Annealing.

Skills

The tools I've used in the projects above.

Machine learning and data
PyTorch, scikit-learn, pandas, NumPy, Albumentations, CNN, RNN and LSTM, NLP, computer vision, transfer learning, regression, cross-validation, hyperparameter tuning, EDA, feature engineering, FastText, Sentence Transformers, semantic search, PCA, TF-IDF
Computer graphics and 3D
Unreal Engine 5, Cesium for Unreal, 3D simulation, astronomical computation (Skyfield, JPL DE441)
Software engineering
Data structures and algorithms, OOP, multithreading, UDP networking, search algorithms (A*, BFS, LRTA*, Simulated Annealing)
Parallel and GPU computing
CUDA, OpenMP, MPI, Pthreads, Intel VTune, Nsight Compute, perf
Programming languages
Python, C++, C, SQL, Java, Kotlin, Bash, HTML
Tools and platforms
Git, GitHub, PyCharm, IntelliJ, Linux, Windows
Spoken languages
Romanian (native), English (B2), French (A2 to B1)

Education

M.Sc. in Data Science

Politehnica Bucharest, Sep 2026 to present

Admitted early, in May 2026.

B.Sc. in Computer Science

Politehnica Bucharest, Oct 2022 to Jul 2026. GPA 3.83/4

Coursework: Data Structures and Algorithms, OOP, Parallel and Distributed Algorithms, Operating Systems, Programming Paradigms, Communication Protocols, Advanced Graphic Programming, Machine Learning.

Saint Sava National College

Sep 2018 to Jun 2022

Recognition

Honorable mention

Scientific Communications Session for Students, May 2026

Presented the eclipse simulator, which I later released as a downloadable product.

Three national prizes, 2021

1st prize, National Physics Contest "Phi". 3rd prize, "Gazeta Matematică" National Olympiad. 3rd prize, National Physics Olympiad.

3DPUB Summer School

Computer Graphics, Jun to Sep 2025

ECDL Standard Certificate

7 modules, 2021

Let's talk

I'm open to junior Machine Learning Engineer, Data Scientist and Software Engineer roles, full-time or part-time. Email is the fastest way to reach me.

dragos.boncan@gmail.com LinkedIn GitHub