Radu G.

Artificial Intelligence

• research and benchmarked SOTA in Computer Vision, NLP, and AI-related for 5+ broad intelligent tasks using PyTorch, scikit, NumPy, open
• conceptualized new algorithms for detection and classification with 90-95% accuracies
• integrated, validated and optimized models pipeline for execution on cloud, Linux & ROS

Data processing

• created a standardized dataset for validation of optical character recognition
• integrated feature engineering and sensor fusion with SLAM and navigation algorithms
• designed and constructed a pipeline for complete generalized data processing methods for object detection, segmentations and classifications

Embedded

• engaged in ML-related design and testing of hardware components
• optimized low-level processing (GPU, DLA, PVA) using advanced libraries (Cuda, VPI, TRT)

Architecture and management

• cooperated using git and R&D for a prototype model methodology
• mentored new employees in ML and data science departments
• identifying and adapting business needs to solutions into the product weekly
• contributed to software planning, management decisions
• helped in business-related responsibilities and miscellaneous tasks (from pr, hr to logistics)
• tested and presented regularly the product in demos to investors and stakeholders

Teamwork and methodology

• cooperated using agile methods with 8-12 members
• planned the project timeline for iterative development
• designed the architecture, behaviour and structure
• documented the process consistently, diagnosis and visualization of results

Statistics and Big Data

• dynamic anomaly detection on time series signal using statistical modelling for +99% accuracy
• processing big data streams using spark, Hadoop, & map reduce concepts on 20+ TB dataset
• deployed SOTA solution as a docker container in a multi-node cluster

Machine Learning and Active Learning

• implemented labelling & diagnosis interface in video classification for a 20x faster process
• trained & fine tuned transfer learning CNN model in Keras for the classification of street infrastructure with 98% accuracy
• innovated with a semi-supervised active learning pipeline for similar performance with 8-24x faster training and 10-80x less data (work presented in conferences)