ÃÛÌÒÊÓƵ

Project 1: Design and Evaluation of a Zero-Trust Multi-Agent Framework for Secure Collaborative AI Systems

• Explore and design end-to-end agent authentication and authorization.
• Evaluate dynamic trust assessment and revocation.
• Assess the need to encode zero-trust rules in a declarative policy language.
• Develop autoscaling policies that spin up extra authentication or revocation nodes under load.
• Validate performance improvements through comprehensive benchmarking and testing.

Project 2: Design and Implementation of a Multi-Agent Framework for Hallucination Filtering and Disinformation Mitigation

• Assess hallucination and disinformation patterns.
• Architect multi-agent pipelines.
• Develop hallucination detection, verification and reinforment agents.
• Optimize agent orchestration and resource management.

Project 3: Optimizing Audio and Video Analytics in Scalable LLM RAG Pipelines

• Develop methods to integrate audio and video data streams into the existing text-based RAG framework, enabling the model to handle and interpret multimedia queries.
• Implement and optimize algorithms for real-time processing of audio and video data within the LLM framework, ensuring efficient retrieval and generation without compromising the response times.
• Enhance the retrieval mechanisms to support multimedia data, including the development of feature extraction techniques and indexing strategies that are optimized for audio and video content.

Project 4: Extending LLM Guard Library to Address the

• Conduct a gap analysis comparing current LLM Guard capabilities against the OWASP Top 10 for LLM applications
• Develop threat models and realistic attack scenarios (e.g., prompt injection, data exfiltration, tool misuse)
• Design and prototype new guardrails, filters, or validation mechanisms to mitigate identified risks
• Implement and test enhancements within the LLM Guard framework
• Evaluate effectiveness using adversarial testing and red-teaming techniques
• Document findings and contribute code, documentation, or guidelines to the open-source community

Project 5: Modeling AI Products for Business Model Classification

• Design and refine an AI model that classifies AI products into the business model archetypes defined in Weber (2022).
• Use commonly reported strategic and operational metrics, such as Resource-Based View (RBV) and Dynamic Capabilities (DC) to inform the model.
• For any metrics that cannot be directly calculated, establish clear and justifiable assumptions, along with guidelines that define how those values should be estimated.

Project 6: Optimization of Indexing and Retrieval Algorithms For Large-Scale Milvus Vector Stores Deployed on GKE

• Implement optimized indexing and retrieval strategies tailored to Milvus vector store operations.
• Refine resource allocation and scaling policies for Milvus deployed on GKE for optimal performance.
• Integrate efficient caching and query optimization techniques.
• Validate performance improvements through comprehensive benchmarking and testing.
• Study the impact of various Retrieval and Indexing algorithms on the overall accuracy, retrieval performance, and resource consumption.
• Build a benchmarking framework characterizing the energy profile of common RAG indexing and retrieval patterns.
• Design guidelines and best practices for building energy-aware RAG pipelines.

Project 7: SLM-driven robotics in MuJoCo with ROS 2 coordination

• Build a working simulation testbed where a Small Language Model handles high-level reasoning for one or more robots running in MuJoCo, with Gazebo-like-ROS 2 backend connecting the pieces.
•  Study how well an SLM, rather than a frontier LLM, can plan, decompose tasks, and re-plan when things go wrong, inside a physics simulator that runs fast enough.
• Build a working pipeline where an SLM under 8B parameters can drive at least one MuJoCo robot through a multi-step task it was given in natural language. 
• Scale the pipeline to several robots running in parallel, coordinated through ROS 2 topics and services. 
• Evaluate the resulted pipeline and revisit the tech stack retroactively to investigate the appropriateness of the selected simulation software and the ROS2.