Sima Didari
Senior Staff AI Scientist · AI Science Lab, Samsung SDS · San Jose, CA
I am a Senior Staff AI Scientist with 10+ years of experience in Deep Learning, Computer Vision, Large Language Models, and Robotics. My work focuses on product development from research to deployment, leading cross-functional teams to deliver AI solutions that create business value.
I contribute to the research community through publications in peer-reviewed venues and hold multiple patents in AI and machine learning technologies. The research tabs below describe each of my areas in more detail, along with the publications that came out of them.
News
Latest research news and publications from this year. A complete list of publications and citations is on my Google Scholar profile.
QAFD-RAG introduces the first principled flow diffusion for graph-based retrieval-augmented generation, reweighting graph edges online according to the meaning of the query. It comes with exponential convergence guarantees and reaches state-of-the-art results across multiple benchmarks. The work is open sourced — code, models, and the paper are all available from the project page.
Research Interests
My research develops methods that advance the state of the art while addressing real-world challenges in industry. Each area below lists its own related publications.
Robotics and Physical AI
I work on data-efficient, reliable vision-language-action (VLA) models for real-world humanoid robot use cases, covering the full path from data collection through model development to deployment. Physical AI is the broader thread: giving embodied systems models that perceive, reason about, and act in the physical world, with the reliability and uncertainty awareness that safe deployment on real hardware demands.
Current Focus
- Vision-language-action models: data-efficient, reliable VLA policies for real-world humanoid robots
- Robot data pipelines: teleoperation and demonstration collection, curation, and scaling strategies for embodied datasets
- Embodied multimodal perception: grounding language and visual observations in 3D scene understanding for manipulation
- Reliability and uncertainty: quantifying policy uncertainty so embodied systems know when to act and when to defer
- Deployment on real hardware: closing the gap between benchmark performance and behavior on physical robots
Generative AI & Large Language Models
My work in generative AI focuses on making language models more reliable and more data-efficient: agentic models and memory management for multi-agent systems (MAS), data-efficient alignment techniques, retrieval-augmented generation, and multimodal LLMs tailored to text, image, and tabular data.
Current Focus
- Agentic models & multi-agent systems: agent design and memory management for MAS
- Data-efficient alignment: preference training and fine-tuning that need less human feedback
- Uncertainty estimation: proxy-based uncertainty for improved instruction following
- Retrieval-augmented generation: graph-based RAG with retrieval guarantees for knowledge-intensive tasks
- Text-to-SQL: reinforcement learning with partial-match and verbal rewards for structured query generation
- Multimodal LLMs: unified handling of text, image, and tabular inputs
Computer Vision
My computer vision research centers on systems that learn effectively with minimal supervision while providing reliable uncertainty estimates. I build active learning and unsupervised accuracy estimation frameworks for classification, object detection, and semantic segmentation, including a probabilistic multi-view 3D semantic segmentation active learning framework.
Current Focus
- Active learning: Bayesian frameworks for classification, object detection, and semantic segmentation
- Unsupervised accuracy estimation: assessing model quality without labeled evaluation data
- Multi-view 3D understanding: probabilistic 3D semantic segmentation across viewpoints
- Representation learning: self-supervised and contrastive methods that cut annotation cost
- Medical and industrial imaging: uncertainty-aware vision for high-stakes applications
Experience
Promoted to Senior Staff AI Scientist in 2024
Key Responsibilities
- Develop and deploy products, solutions, and service-differentiating technologies to secure Samsung's competitive advantages
- Establish technical goals, lead task forces, and collaborate across Samsung and with external partners on technology development and transfer
Major Projects
- Robotics: Developed data-efficient, reliable vision-language-action (VLA) models, covering data collection, model development, and deployment for real-world humanoid robot use cases
- Generative AI & Large Language Models: Developed agentic models, memory management for multi-agent systems (MAS), data-efficient LLM alignment techniques, retrieval-augmented generation (RAG), and tailored multimodal LLMs for text, image, and tabular data
- Computer Vision: Developed active learning and unsupervised accuracy estimation frameworks for classification, object detection, and semantic segmentation; created a probabilistic multi-view 3D semantic segmentation active learning framework
- Developed and deployed AI solutions to enhance engineering, services, and supply chain performance
- Led AI application productization in collaboration with software, UI, and DevOps teams, as well as internal and external customers
- Created end-of-life prediction Deep Learning and Machine Learning models for SmartFactory service solutions
- Deployed a real-time Deep Learning anomaly detection web application for a semiconductor smart manufacturing platform
- Developed a Gaussian Process Machine Learning framework for modeling the impact of process and equipment parameters on sputtering systems (in collaboration with MIT)
- Developed probabilistic methods and Machine Learning models for calibration, validation, uncertainty quantification, optimization, and meta-modeling of asset digital twins
- Contributed to GE's internal Gaussian Process Python package and Uncertainty Quantification application
- Developed Bayesian Machine Learning solutions for life cycle prediction, design space optimization, and supply chain management
- Built real-time defect recognition and image processing software for 3D CT composite data
- Conducted research in computer-aided engineering (CAE), topology optimization, and image processing
- Performed topological characterization and optimization of 3D composite porous microstructures
Education
Patents
Nine granted US patents spanning machine learning, semiconductor manufacturing, and bioprocess systems.
AI & Machine Learning — Samsung SDS America
Semiconductor Manufacturing — Applied Materials
Physical AI — GE / Global Life Sciences Solutions
Awards & Recognition
One of 50 employees selected from across all divisions of Samsung SDS worldwide for outstanding performance; participated in special training on the business and culture of Samsung SDS.
Granted by the CEO for outstanding performance, creativity, organizational ability, and teamwork.
For “Medical Image Labeling via Active Learning is 90% Effective.”
Development of an automated probabilistic testing framework.
Outstanding contribution to the Artificial Lift Optimization project.
Outstanding contribution to Assisted Defect Recognition tools for composite microstructures.
International Mechanical Engineering Congress and Exposition.