Sadegh Keshtkar
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
An Agentic Retrieval-Augmented Generation Framework for Modular Knowledge PipelinesApply
This research investigates an agentic Retrieval-Augmented Generation (RAG) framework in which retrieval, indexing, data preprocessing, and knowledge extraction are performed by explicitly defined and composable agents. Each agent is responsible for a specific phase of the RAG pipeline and can be dynamically selected, replaced, or orchestrated based on task requirements and domain constraints. The proposed system enables users to inject domain-specific logic into retrieval and indexing processes, moving beyond static, monolithic RAG architectures. The work evaluates the effectiveness, flexibility, and performance trade-offs of agent-based pipelines compared to conventional RAG systems.
A Collaborative Human–Agent RAG Chat System with Role-Aware Reasoning and Expert-Owned KnowledgeApply
This thesis proposes a collaborative RAG-based chat system that integrates end users, autonomous language model agents, and human domain experts within a single conversational environment. Each participant operates under a distinct role with different priorities, permissions, and perspectives over the shared conversation state. The system supports structured reasoning, planning, evaluation, and response generation, while enabling experts to maintain personalized RAG representations derived from their historical interactions and domain knowledge. Additionally, the system leverages conversational interactions to generate high-quality question–answer pairs as auxiliary training data for improving retrieval and knowledge grounding.
Regulation-Aware AI Supervision: A RAG-Based Evaluation and Policy Enforcement FrameworkApply
This research proposes a regulation-aware AI supervision framework that evaluates, filters, and constrains the inputs and outputs of AI systems to ensure compliance with legal, ethical, and operational requirements. Using a Retrieval-Augmented Generation (RAG) approach, the system dynamically retrieves applicable regulations, standards, and policies and incorporates them into the decision-making and validation process. The framework supports modular policy agents that can act as evaluators, validators, or content filters, enabling region- and domain-specific governance of AI behavior.
AgentFlow: A Modular Pipeline for Coordinated Collaboration of AI AgentsApply
AgentFlow is a modular orchestration framework designed to coordinate multiple specialized AI agents for complex, multi-stage tasks across heterogeneous data modalities. The system enables structured collaboration among agents such as planners, generators, evaluators, and tool executors, allowing dynamic transitions between stages including extraction, reasoning, transformation, and output generation. AgentFlow supports hierarchical workflows, dependency management, and adaptive agent selection based on task requirements.
Advanced Retrieval-Augmented Generation: Improving Quality, Latency, and AdaptabilityApply
This research explores advanced Retrieval-Augmented Generation (RAG) architectures aimed at improving response quality, reducing end-to-end latency, and enhancing adaptability across domains. The work investigates techniques such as hybrid retrieval, dynamic query rewriting, contextual re-ranking, and distributed indexing strategies. The proposed system adapts retrieval and generation behavior based on domain characteristics and workload constraints, enabling scalable and context-aware AI systems.
Federated Fine-Tuning of Large Language Models in Distributed and Privacy-Preserving EnvironmentsApply
This research investigates federated learning approaches for fine-tuning large language models (LLMs) across distributed environments without centralizing data. The framework enables collaborative model improvement while preserving data privacy and security. Key challenges such as data heterogeneity, communication efficiency, system scalability, and convergence stability are addressed through adaptive aggregation, selective parameter updates, and model distillation techniques.
Agentic Retrieval-Augmented Generation System for Modular Knowledge PipelinesApply
This research proposes an agentic Retrieval-Augmented Generation (RAG) system in which retrieval, indexing, data preprocessing, and knowledge extraction are handled by explicitly defined and composable agents. Each agent encapsulates domain-specific logic and can be dynamically selected or orchestrated within a pipeline. The approach enables flexible, transparent, and reusable knowledge workflows, moving beyond static RAG architectures toward user-extensible retrieval systems.
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 partitioning the model into segments that are trained sequentially. Only a subset of segments is loaded into memory at any given time, enabling training on resource-constrained hardware. The work addresses challenges such as inter-segment dependency management, gradient consistency, and efficient backpropagation across unloaded components.
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), 2024-11-26 BibTeX URL DOI
All publications as BibTex