Open to opportunities

Roman Talyansky

@romantalyansky

Message

Applied AI researcher and research engineer focused on retrieval, NLP, multimodal AI, and scalable LLM systems

Israel

Message

What I'm looking for

I’m looking for senior applied AI research or research engineering roles focused on search, retrieval, NLP, multimodal AI, and LLM systems. I enjoy building production-oriented AI systems that balance model quality with latency, scalability, and cost, and value technical ownership and end-to-end impact.

Applied AI researcher and research engineer with experience spanning search, retrieval, NLP, recommendation systems, multimodal AI, and large-scale distributed ML infrastructure.

My recent work focuses on multimodal search systems, including retrieval, reranking, and LLM-based query refinement, as well as LLM-integrated retrieval workflows. I have designed and implemented end-to-end prototypes that combine model capabilities with practical system considerations such as latency, throughput, and scalability. I enjoy working at the intersection of model design and system design, where improving overall system behavior requires both.

In parallel, I have extensive experience in large-scale training and optimization of deep learning systems. This includes work on distributed training, data and pipeline parallelism, and efficiency improvements for both recommendation models and large language models. My contributions include optimization methods that improved scalability and training efficiency, as well as work integrated into production-grade frameworks such as NVIDIA Megatron.

Earlier in my career, I worked on information retrieval and recommendation systems, building search and data-mining capabilities over large-scale enterprise data. I also developed NLP-based systems for automatic answering of customer messages, summarization, topic segmentation, named entity recognition, and sentiment analysis. These experiences gave me a strong foundation in language and retrieval systems prior to the current wave of LLM-based approaches.

Across roles, I have consistently taken applied research from idea through implementation, evaluation, and deployment, working closely with engineering teams to deliver production-impact systems. I am particularly interested in building AI systems that combine strong model capabilities with efficient, scalable infrastructure, especially in areas such as search, retrieval, and AI systems interacting with real-world data.

Experience

Work history, roles, and key accomplishments

Current

Senior AI Researcher

Current

Toga Networks (Huawei Research)

Mar 2025 - Present (1 year 2 months)

Designed and implemented an end-to-end multimodal search proof of concept, including retrieval, reranking, and LLM-based query refinement in an LLM-integrated retrieval workflow. Built a pipeline-parallel strategy for heterogeneous execution achieving ~9% higher throughput vs a strong data-parallel baseline and evaluated quantization approaches to improve training efficiency.

Multimodal Search LLM Based Query Refinement Information Retrieval Pipeline Parallelism PyTorch C++

Senior AI Researcher

Toga Networks (Huawei Research)

Aug 2018 - Mar 2025 (6 years 7 months)

Led applied research on scalable production AI systems focused on efficiency, throughput, and system utilization at large scale. Developed optimizations for distributed DLRM training increasing feasible batch size several-fold beyond MLPerf baselines, and built an asynchronous distributed optimizer for LLM pipeline model-parallel systems integrated into NVIDIA Megatron to reduce training time whil

Distributed Training Asynchronous Distributed Optimization Pipeline Parallelism NVIDIA Megatron Large Scale ML Infrastructure Throughput Convergence & Accuracy

Senior Machine Learning Researcher

Amazon Research

Jul 2017 - Aug 2018 (1 year 1 month)

Led applied ML research at Amazon Go to reduce human-in-the-loop annotation, improving operational efficiency and contributing to business performance through production deployment. Built and tuned real-world AI systems balancing model quality, automation, and operational cost under production constraints.

Production AI Systems AI Systems Model Quality Human In The Loop Deployment Focused Evaluation Applied Research

Senior Researcher

Toga Networks (Huawei Research)

May 2014 - Jul 2017 (3 years 2 months)

Led development of DeepSpark, an asynchronous distributed system for training deep learning models on Apache Spark clusters. Developed asynchronous training algorithms and storage-less fault tolerance for large-scale ML systems to improve robustness and training scalability.

Apache Spark Distributed Training DeepSpark Python C++ Prototyping Reliability Engineering

Researcher

SAP Research

Aug 2008 - May 2014 (5 years 9 months)

Designed and implemented information retrieval capabilities on SAP HANA to enable efficient search over large-scale enterprise data. Built data-mining solutions for a context-aware recommendation system and developed NLP methods for sentiment analysis in low-label settings, including keyword extraction and tag generation.

Information Retrieval SAP Hana Recommendation Systems Sentiment Analysis Low Label NLP Keyword Extraction Data Mining Entity Text Processing Statistical Methods

Senior Architect

Semantica

Dec 2007 - Aug 2008 (8 months)

Led NLP and ML development for automatic analysis and answering of customer SMS and email messages. Delivered end-to-end implementation from concept through deployment for customer communication understanding.

NLP SMS Email Processing Model Evaluation

Architect

Intel

May 2002 - Dec 2007 (5 years 7 months)

Developed a natural-language spoken interface prototype and designed extensions to binary instrumentation frameworks. Worked on prototype and engineering efforts to enable language-driven interaction and instrumentation capabilities.

Speech NL Prototyping Systems Engineering Interface Design Optimization

Senior Researcher

Siftology

Oct 2000 - May 2002 (1 year 7 months)

Developed NLP algorithms for summarization, topic segmentation, and named entity recognition. Deployed multilingual NLP systems across multiple languages to support production text processing workflows.

Summarization Segmentation Named Entity Recognition Evaluation Pipelines

Researcher

IBM Research Lab

Aug 1997 - Oct 2000 (3 years 2 months)

Contributed to the design and development of GPFS distributed file system, including optimizations for disk load balancing and quota management. Worked on core distributed systems components supporting reliable high-performance storage.

Quota Management Storage Systems Distributed Systems Performance Optimization