Sadegh Keshtkar
sadegh.keshtkar@gwdg.de |
Biography
Since December 15, 2022, Mr. Sadegh Keshtkar has been an AI Developer in the “Computing” working group (AG C). He completed his second Master's in Computer Science at Tu Berlin and is now pursuing a Ph.D. in Computer Science at GWDG, under the supervision of Prof. Julian Kunkel. At Tu Berlin, his Master's thesis focused on studying how an automated red teaming agent learns using deep reinforcement learning. For his Ph.D., he's diving into federated learning's characteristics, exploring ways to improve this area of study with Prof. Kunkel's guidance.
ORCID: 0000-0002-4683-0136
Research Interests
- Federated learning
- Reinforcement learning
- Application of machine learning methods
Teaching
Open Thesis Topics
Regulation-Aware AI Supervision: RAG-Based Evaluation and Filtering FrameworkApply
This project proposes an AI-driven supervision system that evaluates, filters, and regulates the input and output of AI services to ensure alignment with legal, ethical, and operational constraints. Leveraging a Retrieval-Augmented Generation (RAG) approach, the system dynamically retrieves applicable rules, standards, or compliance requirements and uses them to guide or constrain AI outputs. The framework can act as an evaluator, validator, or content filter, and is capable of binding model behavior to regional or domain-specific regulations through modular policy enforcement agents.
AgentFlow: A Modular Pipeline for Coordinated AI Agent CollaborationApply
AgentFlow is a system designed to orchestrate multiple specialized AI agents for handling complex, multi-stage tasks across diverse data modalities. It enables structured coordination among agents such as generators, evaluators, planners, and tool-runners, allowing for seamless transitions between steps like data extraction, reasoning, transformation, and final output generation. The system supports hierarchical workflows and dynamic agent selection based on task needs, improving both automation and adaptability.
Advanced Retrieval-Augmented Generation (RAG): Enhancing Quality, Speed, and AdaptabilityApply
This research explores the next generation of Retrieval-Augmented Generation (RAG) systems with a focus on improving response quality, reducing latency, and optimizing both indexing and retrieval performance. It integrates advanced re-ranking, dynamic rewriting, and hybrid search techniques to better align the retrieved context with the user query. The system also adapts to domain-specific needs by incorporating fast, distributed retrieval pipelines and context-aware generation.
Federated Fine-Tuning of Large Language Models (LLMs) in Distributed EnvironmentsApply
This topic investigates the design and implementation of a federated learning framework for fine-tuning large language models across distributed and privacy-preserving environments. The system supports collaborative learning without centralizing data, enabling secure and scalable model updates. Key challenges such as heterogeneity, communication overhead, and convergence efficiency are addressed through intelligent orchestration, model distillation, and adaptive optimization strategies.
On-Device Federated Training with ONNX Runtime: A Solution for Ecosystem HeterogeneityApply
This research explores a federated learning framework utilizing ONNX Runtime APIs to enable efficient, on-device training of AI models across heterogeneous platforms. The system addresses the diversity in hardware, operating systems, and model formats by leveraging the portability and interoperability of ONNX. It ensures privacy-preserving learning while managing challenges such as inconsistent compute resources, non-IID data distribution, and communication constraints. The goal is to build a flexible and scalable solution for federated training in real-world, multi-device AI ecosystems.
Segment-Wise Sequential Fine-Tuning of Large Language Models Under Memory ConstraintsApply
This research investigates a memory-efficient fine-tuning strategy for large language models (LLMs) by splitting the model into segments and training them sequentially. Only a subset of model segments is loaded into memory at any given time, enabling fine-tuning on resource-constrained devices. The project addresses key challenges including segment dependency management (parallel vs. sequential paths), efficient backpropagation across unloaded segments, and maintaining gradient consistency. This approach opens new possibilities for LLM training without requiring full-model memory allocation.
Theses
- Learning to Attack: Automated Red Teaming Using Deep Reinforcement Learning, Sadegh Keshtkar (Master's Thesis), Advisors: Sahin Albayrak, 2022-05, BibTeX
Publications
2024
- State-of-the-art artificial intelligence techniques in healthcare publications, and their correlation with disease and data: A data driven analysis (Sadegh Keshtkar, Dagmar Krefting, Anne-Christin Hauschild, Zully Maritza Ritter, Narges Lux, Aasish Kumar Sharma, Pavan Kumar Siligam, Julian Kunkel), In Journal of Artificial Intelligence and Robotics (1) \myDOI{10.52768/JArtifIntellRobot/1014}, 2024-11-26 BibTeX URL DOI
All publications as BibTex