Specialized AI, built around the problem
My experience combines neural model development with the engineering needed to make models useful. Across production work and competition projects, I have worked with language, images, speech, and structured data—adapting established methods to specific tasks and constraints.
Neural machine translation
I independently developed the translation core within the team-built Linguist platform. This work includes recurrent and Transformer-based models, language-specific preprocessing, postprocessing, and document-format handling across 64 translation directions. The engineering challenge extends from preparing language data to preserving the structure of translated business documents.
Methods and tools: RNN-based NMT, Transformers, OpenNMT, Marian, language processing pipelines.
Computer vision and segmentation
My project and competition experience includes medical images, product classification, camera identification, wildlife monitoring, and image segmentation. I have worked with CNN families including VGG, ResNet, DenseNet, Inception, Xception, and InceptionResNetV2, as well as U-Net segmentation models. The work combines model adaptation, data preparation, augmentation, ensembles, and classical image processing.
Methods and tools: CNNs, U-Net, transfer learning, ensembles, TensorFlow, Keras, OpenCV.
Speech and sequence modelling
My speech recognition work has combined Mel-spectrogram and MFCC features with LSTM models. Broader competition experience includes biological signals, video understanding, demand forecasting, and other prediction tasks where the structure of the data matters as much as the model.
Methods and tools: LSTM, recurrent models, signal preprocessing, feature engineering.
Connected AI systems
I also build agents that work with company databases, knowledge bases, Jira, GitHub, and AWS. Selected client implementations support infrastructure analysis, bottleneck identification and remediation, feature assessment, and security review. I connect these capabilities to the systems and decisions they are intended to support.
Tested across domains
I’m a Kaggle Competitions Master with nearly 100 competitions and 18 medals: one gold, two silver, and fifteen bronze. This experience has exposed me to different data types, evaluation methods, and domain constraints, helping me get productive quickly when entering a new field.
From model to working system
I start with the outcome, available data, and practical constraints. I assess whether the task needs a specialized model, an existing model, or a simpler approach, then define evaluation criteria and the path to integration. You work directly with the person making those decisions and building the solution.