stanislau beliasau

research engineer, member of technical staff

• contributed to mirage, a 13B-parameter text+audio+image-to-video model. built distributed training on 1000+ H100 GPUs at 40% MFU; context-parallel up to 150k tokens.
• optimized mirage video inference: 4x faster startup, 2x overall speedup via fp8 quantization, custom attention kernels, torch compile, and inference caching.
• optimized mirage audio (10B text+audio-to-audio): +30% training throughput, 4x inference speedup via efficient kernels, torch compile, and HSDP.
• designed and implemented a novel audio-to-landmarks architecture for lip-sync.
• built an automated eval framework for continuous checkpoint assessment (pub/sub, W&B logging, cloud uploads).
• evaluated GPU providers and drove infra decisions for efficient multi-node training at scale.

ml course tutor

• gave lectures on NLP and computer vision; designed coursework on distributed LLM training (GPT-2XL, OPT-6.7B).

computer vision engineer

• designed novel image-encoding for personalized stable diffusion, replacing dreambooth fine-tuning with single-shot encoding (predated similar 2024 publications).
• trained real-time mobile GANs at HD/60fps with sub-2MB model size.
• 30% stable-diffusion inference speedup via architectural changes and efficient sampling.
• shipped production models for re-aging, body reshaping (image+video), body segmentation.
• deployed to iOS via CoreML and server via TorchScript.

machine learning engineer

• built a gRPC + FastAPI image-classification service with Uvicorn, Prometheus, Docker, Supervisord.
• custom architectures in C++, ~45% latency reduction.
• synthetic data improved accuracy by up to 10%.
• distilled a CNN into a 3x smaller model with near-identical metrics.
• time-series demand forecasting reduced MAPE by 20%; AirFlow production pipelines.

software engineer intern

• rule-based + NLP solutions for affiliations parsing.
• flask-based annotation service, dockerized and deployed.
• similarity metrics for log search.

master's-level ml academic program

• coursework: efficient DL systems, RL, computer vision, NLP, recommender systems.

bsc, computer science

• full government scholarship awarded on entrance-exam results.