Verified Expert in Engineering
Artificial Intelligence Developer
Andrei studied computer science in his hometown in Romania and completed his master's degree in AI at the University of Amsterdam. He has accumulated practical experience in AI over years of training, developing data processing pipelines, and deployment. He is an engineer always looking forward to new challenges. Andrei also has academic experience through publishing two papers on neural pruning and quantization, which were well received by the academic community.
Linux, Visual Studio Code (VS Code), Slack, Rocket.Chat, Google Cloud/Suite
The most amazing...
...thing I've accomplished is earning the Best Paper award for publishing my master's dissertation in BeneLearn 2020 about a novel pruning algorithm I developed.
Machine Learning Engineer
- Built a custom and flexible BERT-like architecture for multi-class document classification and trained on data from various clients, obtaining 90-94% average accuracy.
- Combined the traditional NLP augmentation model with the GPT-3 large language model to do data augmentation for clients, reducing the error rate by 50%.
- Used Hugging Face datasets based on Apache Arrow to handle large volumes of data that normally would not fit in memory and implemented an efficient and replicable data processing pipeline with batching and multiprocessing.
- Held regular client meetings, giving high-level overviews of our technical solution and explaining our core metrics.
Machine Learning Engineer
- Trained a YOLOv5 object detection network for waste management and recycling, obtaining mean-average-precision scores of over 95% on over 40 classes of objects with a speed of over 200 frames per second on a conventional GPU.
- Built an app around the object detection network using FastAPI to expose the endpoints and Streamlit for the UI, converting the network to the ONNX format after training for faster inference time.
- Created a module for out-of-distribution detection using the CLIP pre-trained model, obtaining over 97% accuracy in the in-/out-of-distribution classification while allowing for the class taxonomy to be changed without the need for re-training.
- Held regular meetings with project stakeholders, led demos and presentations for the client to provide estimates for the next milestones, and conducted sprint planning sessions, keeping track of progress.
- Published the paper: Highlights of AI Research in Europe in a special edition of the European Journal of AI, demonstrating that pruning and quantization can bring greater amounts of acceleration when used in conjunction without sacrificing accuracy.
Machine Learning Research Internship
- Researched and built expertise in neural network pruning techniques as part of my master's dissertation.
- Developed a novel pruning algorithm that obtains state-of-the-art results for high sparsity scenarios and other properties such as the ability to prune during training, computational tractability, and hyperparameter invariance.
- Received the Best Paper award for writing a scientific paper around said algorithm and publishing it at the BeneLearn 2020 conference held in Belgium, Netherlands, and Luxembourg.
- Built a time series forecasting model with the SARIMA method, achieving a low mean square error for all predictions within the confidence bound.
- Created an encoder/decoder gated recurrent unit network for document part classification, obtaining over 90% accuracy.
- Deployed the trained models to production by exposing the core functionality via RESTful APIs and monitored the performance in production.
Research Scientist Intern
- Analyzed customer behavior on the platform and developed a random forest model with a high ROC-AUC score.
- Handled large volumes of data using Apache Spark and created data processing pipelines to filter and prepare data using the Python and Scala APIs and Spark SQL.
- Conducted A/B testing and integrated the resulting model into several Amazon sites.
FlipOut | Uncovering Redundant Weights via Sign Flippinghttps://github.com/AndreiXYZ/flipout
It can remove over 98% of the connections in common networks with little to no impact on accuracy, allowing for large speed gains. Compared to baselines from literature, this method can prune during training, is insensitive to the selection of hyperparameters, and allows for selecting the sparsity level directly.
I wrote a paper around this method and published it in BeneLearn 2020, obtaining the Best Paper award.
Python 3, Python, Bash, SQL
Flask, Apache Spark
PyTorch, Matplotlib, OpenCV, Scikit-learn, Pandas, PyTorch Lightning, NumPy
TensorBoard, You Only Look Once (YOLO), Slack, GitHub, Git, Docker Compose, Spark SQL
Visual Studio Code (VS Code), Rocket.Chat, Linux, Docker, Zeppelin
Machine Learning, Deep Learning, Natural Language Processing (NLP), Computer Vision, Long Short-term Memory (LSTM), Deep Neural Networks, Neural Network Pruning, Quantization, FastAPI, Object Detection, Transformers, BERT, Classification, English, Google Cloud/Suite, Artificial Intelligence (AI), Proof of Concept (POC), Minimum Viable Product (MVP), Natural Language Understanding (NLU), Contract, Information Extraction, GPT, Generative Pre-trained Transformers (GPT), Statistics, Software Engineering, Scientific Data Analysis, Random Forests, APIs, Scientific Computing, Research, NVIDIA Triton, Streamlit, Detection Engineering, Open Neural Network Exchange (ONNX), Hugging Face, DVC, Machine Learning Operations (MLOps), Language Models, Algorithms, Data Analytics, Large Language Model (LLM), TensorRT, Image Processing, Generative Adversarial Networks (GANs), Generative Artificial Intelligence (GenAI), OCR, PDF, Information Retrieval, Cluster Computing, Time Series, ARIMA Models, DeepSpeed, 3D
Object-oriented Programming (OOP), Data Science, Agile, REST
Master's Degree in Artificial Intelligence
University of Amsterdam - Amsterdam, The Netherlands
Bachelor's Degree in Computer Science
Alexandru Ioan Cuza University - Iași, Romania
Model Parallelism: Building and Deploying Large Neural Networks
IELTS Academic Certificate (Native Level)