At SoftConstruct, I’m building a real-time, pose-driven AI live avatar for casino games, combining full-body video generation with motion control and targeting inference on a single H100.
I fine-tuned video diffusion and flow-matching models, built a synthetic data pipeline using Wan-Animate, and developed a markerless motion capture system for full-body and finger motion. I’m also working on streaming distillation and low-latency inference.
At Metric, I developed vision-language models, Armenian text embedding models, LLM agent chatbots, and RAG pipelines. I also developed a real-time multi-camera system for vehicle detection, tracking, and traffic-violation analysis, and optimized models for NVIDIA Jetson devices.
At Pixeria, I trained diffusion models for room-image generation and applied GANs, VAEs, and style transfer to image generation. My work also included image similarity, classification, and classical image processing.
