cv

Experience

Research Engineer, Member of Technical Staff July, 2023 - Present

Contributed to development of Mirage, a 13B parameter text+audio+image-to-video generation model. Designed novel architectural components, implemented distributed training pipeline on 1000+ H100 GPUs achieving 40% MFU, and developed context-parallel training for sequence lengths up to 150k tokens.
Optimized Mirage Video inference achieving 4x faster startup times and 2x overall speedup through fp8 quantization, custom attention kernels, torch compilation, and inference-time caching with minimal quality degradation.
Optimized Mirage Audio model, a 10B parameter text+audio-to-audio model, achieving 30% training throughput improvement and 4x inference speedup through efficient kernels, torch compilation, and HSDP while maintaining quality.
Designed and implemented novel audio-to-landmarks architecture for lip-sync generation through extensive architecture exploration and ablation studies.
Designed and implemented automated evaluation framework for continuous checkpoint assessment using pub/sub architecture, integrating metrics computation, Weights&Biases logging, and automated video artifact uploads to cloud storage.
Evaluated GPU providers and made key infrastructure decisions enabling efficient multi-node training at scale.

ML course tutor assistant February, 2022 - June, 2023

Served as ML course tutor at Yandex School of Data Analysis, giving lectures on NLP and computer vision, and designing coursework on distributed LLM training (GPT-2XL, OPT-6.7B).

Computer Vision Engineer August, 2022 - May, 2023

Designed and implemented novel image encoding method for personalized Stable Diffusion generation, replacing expensive DreamBooth fine-tuning with single-shot encoding from few images. Approach enabled identity-preserving generation without model fine-tuning; similar methods were later published and widely adopted in 2024.
Developed real-time GANs running on mobile devices at HD quality, 60fps while maintaining sub-2MB model size through aggressive quantization and architecture optimization.
Optimized Stable Diffusion inference achieving 30% speedup through architectural modifications, custom modules, and efficient sampling strategies.
Trained production models for re-aging, body reshaping (images and video), and body segmentation, iterating rapidly on novel architectures and dataset curation strategies.
Deployed models to iOS using CoreML and server infrastructure using TorchScript, optimizing for both on-device and cloud inference.

Machine Learning Engineer February, 2022 - August, 2022

Designed and implemented an image classification service using the gRPC endpoint client/server architecture and the FastAPI framework.
Utilized Uvicorn and Prometheus in conjunction with Docker and Supervisord to create a robust and scalable solution.
Developed and implemented custom model architectures using C++, resulting in up to a 45\% reduction in model latency.
Generated synthetic datasets to supplement real data, leading to an increase in model accuracy of up to 10%.
Successfully distilled the CNN model into a model that was 3 times smaller while maintaining nearly identical evaluation metrics.

Created and curated custom datasets from unstructured client data using Pandas and SQL.
Trained numerous time-series models for demand forecasting, reducing forecast MAPE by 20%.
Constructed production pipelines with AirFlow to convert raw data into a feature vector, feed it into the model, and forecast the product demand.

Software Engineer Intern May, 2021 - November, 2021

Developed rule-based and NLP-based solutions for affiliations parsing.
Developed data annotation service with Flask framework. Wrapped it into Docker and deployed to the server.
Developed similarity metrics for searhing similar logs.

Seeing Voices: Generating A-Roll Video from Audio with Mirage - A. Sundararaman, A. Adishesha, A. Jaegle, D. Bigioi, H. Song, J. Kyl, J. Mao, K. Lan, M. Komeili, S. Athar, S. Babayan, S. Beliasau, W. Buchwalter. arXiv:2506.08279, 2025. [link]
How to Fine-Tune Very Large Model if It Doesn’t Fit on Your GPU - Technical article on distributed training techniques for large language models. 600+ reactions, 2022. [link]

Master's degree level Machine Learning developer academic program September, 2020 - June, 2022

Relevant coursework: Efficient Deep Learning Systems, Reinforcement Learning, Computer Vision, NLP, Recommendation Systems.

Bachelor of Computer Science September, 2018 - August, 2022

Awarded a full scholarship and stipend by the government per entrance exam results.