Portfolio

Enhancing Log Analytics with Generative AI

This thesis explores the integration of Generative AI (GenAI) models, specifically Large Language Models (LLMs), into the field of log analytics to address the growing challenges posed by the increasing volume and complexity of log data in modern IT systems. Traditional log analysis methods, which often rely on manual inspection or rule-based approaches, struggle to keep up with the scale and variability of logs in large-scale environments. To address these limitations, this work introduces the ModuGPT framework, a modular system designed to simplify the development of LLM-powered applications for log analytics. The framework introduces Prompt Elements, a structured catalog of reusable components that streamline prompt engineering and improve the consistency and effectiveness of LLM interactions. Additionally, three specialized tools are developed within the framework: Insight-to-Text, which generates natural language explanations and recommendations for log insights; Text-to-ES, which provides a natural language interface for querying Elasticsearch; and SmartSearch, which enhances traditional search functionality by combining semantic and keyword-based methods.

Efficient Computation of Worst-Case Delay-Bounds for Time-Sensitive Networks

Time-sensitive networks, as in the context of IEEE Time-Sensitive Networking (TSN) and IETF Deterministic Networking (DetNet), require bounds on the worst-case delays as they support safety-critical applications and offer deterministic services with guaranteed, bounded latency. Finding the exact worst-case delays is known to be an NP-hard problem; hence we are interested in bounds on the worst-case delays.

In summary, our project focuses on addressing the challenges associated with the selection of cuts in the PLP algorithm for accurately estimating worst-case delay bounds. We aim to investigate and understand the intricate relationship between the size, shape, and composition of cuts, and their impact on the resulting delay bounds. By gaining insights into this relationship, we can develop more efficient and accurate heuristics for selecting good cuts for networks of varying sizes and topologies. Through these contributions, we strive to improve the reliability and efficiency of communication systems operating in time-critical environments.

Distributed Algorithms Term Project

I implemented perfect links, uniform reliable broadcast and lattice agreement using UDP in Java. My implementation achieved 3M+ message throughput on perfect links level, obtaining the best result in the class.

Concurrent Algorithms Term Project

I implemented software transactional memory using the TL2 algorithm in C. My implementation achieved x2.918 speed up compared to the reference solution using a single global lock, obtaining the maximum grade.

Tolerant Max Cut Tester

We present a sublinear-time algorithm for testing whether a bounded degree expander graph has a max-cut value close to 1 or far from 1. Graphs are represented by incidence lists of bounded length d, and the testing algorithm can perform queries of the form: “who is the i-th neighbor of vertex v”. The tester should determine with high probability whether the max-cut value is greater than 1 − ϵ1 or less than 1 − ϵ2. Our testing algorithm has query complexity and running time O(√n log2 n/ϵ1) where n is the number of graph vertices and ϵ1 is the tolerance value.

Machine Learning Term Project

Infectious and parasitic diseases are difficult to distinguish when the medical equipment is limited. Late detection of these diseases may cause an outbreak and endanger many lives. In collaboration with iGH lab and MSF, we built an unsupervised machine learning model to cluster the patient data to help detect potential outbreaks and their characteristics and a Power BI dashboard for visualization of data and the model.

Computer Networks Term Project

We developed a file transfer client and server application in C#. The client can upload and download files, make the files public or private, make copies on the server, request a list of all files and access other people’s public files. Server can handle multiple client connections at a time. We extensively used the Socket and Threading libraries.

Application of autoencoder neural networks for CFD Problems

Drag and lift forces are important quantities in systems where fluid dynamics are present. Drag force can cause energy losses in systems such as airplanes and trucks. Our goal was to reduce the energy losses caused by drag force. In order to achieve this goal, we considered several approaches.

We managed to accurately predict drag and lift coefficients and reconstruct the encoded images. This model can be used to predict the drag and lift coefficients acting on the system and minimize the drag force acting on the system by optimizing the control parameter.

Algorithms Term Project

Traversing a (finite) directed or undirected graph by passing through each vertex exactly once results in a Hamiltonian Path. Hamiltonian Path Problem is an NP-Complete problem. We analyzed a randomized approximation algorithm both theoratically and experimentally.

GUI Design for Automated Mission Planning and Controller Synthesis

Mission planning and controller synthesis for autonomous robots is a complex task and requires knowledge about several topics such as control, robotics and linear temporal logic. Wrapping these tasks with a graphical user interface and making these tools usable by people without the technical knowledge is important. Because that would result in an increasing number of people getting involved and thus making the field grow.

Motion Control of a Flying Robot (Quadrotor)

I designed a robust controller for hovering and trajectory-tracking of a quadrotor in MATLAB and tested the controller’s ability to perform several different tasks (trajectories) in Simulink.