Alexander Tsvyashchenko

Journal of Machine Learning Research 24(377):1–8

Journal of Machine Learning Research 24(240):1–113

EP4713833A1, Google LLC; with Andrea Gesmundo; as Oleksandr Tsvyashchenko

AI Basics, an online course by Google Ukraine with the Ministry of Digital Transformation of Ukraine

Hack’n’Lead hackathon, Zürich

BotConf, Nancy

Google DeepMind, Zürich, Switzerland

• Research on sparsity, modularity and memory, within internal “frontier challenges” projects that tackle the core research questions for frontier AI models:
  • Lead of the inference sub-stream of the modular, composable and efficient LLM memory project.
  • Co-lead of the modularity in the music domain track.
  • Sparsity in text diffusion.
• Every project is a collaboration with other teams across Google DeepMind.
• Domains: LLM research: modularity, compositionality, sparsity, memory, continual learning, interpretability.
• Skills: Gemini and Gemma training, evaluation and serving stacks, XManager, TPUs, agentic R&D tooling (Antigravity), JAX, Grain, Python, C++.

Google / Brain, Zürich, Switzerland

• Research and development of large-scale ML systems:
  • ML Pathways, T5X and PaLM; co-author of the T5X and PaLM papers (see “Publications”).
  • Multitask and transfer learning for text, vision and multimodal tasks, including applied research in YouTube; the patent on extending multi-task networks to new modalities comes from this work.
  • Sparsity and modularity research, continued at Google DeepMind.
• Host and mentor of interns through the Google internship program.
• Skills: JAX, Flax, T5X, XLA, TPUs, TensorFlow, Python, C++.

Daedalean AI, Zürich, Switzerland

• The research on the visual traffic detection system for uncooperative airborne obstacles: surveyed ML object detection and tracking, then led the development of several approaches for adding temporal information to the detection models.
• Co-designed, with the Simulation and Data teams, the pipeline that models and generates aircraft collision encounters at volume, the camera imaging simulation that narrows the gap to real data, and the traceable data flows that certification requires.
• Designed and implemented the training-to-production model conversion and inference stack for flight-compatible hardware, with proofs of concept for ML kernel verification under DO-178 and for FPGA inference on the Versatile Tensor Accelerator.
• Set up distributed training on Determined AI and optimised the training pipeline.
• Domains: Detect and Avoid; ML object detection and tracking; ML inference, compilers and optimisation; certification and formal verification; embedded devices, GPUs and FPGAs.
• Skills: TensorFlow / Keras, Determined AI, Unigine, GCP, OpenCV; TVM, ONNX / ONNX Runtime, LLVM / Clang, Z3, Nix, Bazel, Yocto, Xilinx Vivado.

FAIRTIQ, Zürich, Switzerland

Advised the research department on sequence models and time series, feature engineering and imbalanced datasets.

Google / Google AI, Zürich, Switzerland

• Owned the full lifecycle of ML projects for ARCore (Tango), YouTube, Abuse and Ads: scoping with the product team, data analysis and collection, model research, evaluation and productionisation.
• Taught the product teams ML along the way, and ML classes inside and outside Google.
• Skills: TensorFlow, TFX; multiple Google technologies, languages and products.

Google / Ad Traffic Quality, Zürich, Switzerland

Traffic analysis, signals, metrics and filters against abusive and fraudulent ad traffic; tech lead of the ad-fraud botnets effort, including information sharing and collaboration with the wider industry (see “Talks”).

Deebmedia, Amsterdam, Netherlands (remote)

Owned the whole video-processing and algorithmic stack; R&D of robust online algorithms for background reconstruction in complex scenes and for highly accurate object tracking.

Atoms Optical Measuring, Locarno, Switzerland (remote)

Researched and prototyped a highly accurate image-segmentation algorithm, one of the key components behind micron-level measurement accuracy.

Automated Industrial Machinery, Inc, Chicago, IL (remote)

End-to-end ownership of four R&D projects for the wire-bending industry: a photogrammetric measurement system, the BenderCad CAD system, real-time bending-machine simulation and CAD wire import; managed three sub-contractors.

Materialise, Kyiv, Ukraine

Researched and implemented the computational-geometry and rapid-prototyping algorithms at the core of Materialise DigitalCAD and of the Magics, Mimics, Simplant and 3-matic products; led a research team of six and introduced automatic testing, autobuild and code-quality processes.

Multiple open-source projects

Stand-alone projects and patches to existing ones, from a Linux kernel Wi-Fi driver to XMPP archiving in Erlang; the recent ones are at github.com/ndl, the older ones under projects.