Explore Projects

<p>I want to improve the Linux kernel environment model for more accurate work with memory.</p>

The future is already here - it's just not evenly distributed” - William Gibson We’d like to introduce an idea for a new open-source project: Racklet. It’s a fully-integrated, Raspberry Pi form-factor server rack and software stack that aims to be a scale model of hyperscaler datacenters. All layers of the stack are 100% OSS/OSH, and will be developed together with the community. It’s reproducible through open PCB designs, 3D printed casing, and commodity, off-the-shelf hardware. We want to lower the barrier of entry for becoming cloud native. Racklet aims to inspire users to explore how modern server architectures work, in a tangible and educational way. Emphasis is put on security, knowledge sharing, extensibility, and portability. The goal is to conceptually map to real environments and provide an accessible and well-documented path to welcome future talents to the world of cloud native.

<p>This project sets out to achieve two goals. The first objective is to update the annotation system for Red Hen’s NewsScape dataset to FrameNet 1.7 using Open-Sesame and Semafor parsers. The second objective is to expand the lexical units, frames and frame-to-frame relations in FrameNet 1.7 through a knowledge-driven approach and a distributional semantics approach. The knowledge-driven approach uses BabelNet to induce the frames of unrecognized lexical units in the tagged NewsScape dataset. The latter distributional semantics approach uses Deep Structured Semantic Models (DSSM) to create word embeddings of lexical units (LUs) to resolve the inconsistency in FrameNet hierarchy, tag LUs with their missing frames, and locate new frames using SemCor corpus. If time permits, DSSM is used to expand the frame-to-frame relations with Entity and Event frames using ACE 2005 Entities and Events dataset.</p>

The purpose of this project is to add Temporal Graph Neural Networks (TGNNs) to GraphNeuralNetworks.jl. The first part of the project will be dedicated to creating a temporal graph type and providing it with all the accessory functions. After that, new layers will be added to the package to build models that fall under the category of STGNNs. Thus, new datasets of temporal graphs will be included in MLDatasets.jl to be used for training the models. Finally, tutorials will be produced to complete the project, introducing the new temporal graph type and showing how to perform temporal graph tasks using the added data, model and features.

<p>Machine learning models are being used to assess loan and job applications, in bail, sentencing and parole decisions, and in an increasing number of impactful decisions. Unfortunately, due to both bias in the training data and training methods, machine learning models can unfairly discriminate against individuals or groups. While there are many statistical methods to alleviate unfairness, there’s a growing awareness that any account of fairness must take causality into account. This project aims to implement causality-based tools for checking and forcing fairness into the Julia ecosystem of machine learning tools. Specifically, we intend to extend the packages Fairness.jl and MLJ.jl to include methods that implement counterfactual fairness, a recent definition of fairness. We will do so building upon principles of causal graphical models and causal probabilistic programming, as implemented in languages such as Omega.jl and CausalInference.jl.</p>

<p>To realise the full potential of probabilistic programming languages (PPLs), it is essential to automate the inference of latent variables in the model, conditioned on the observed data. Hence PPLs usually come with built-in algorithms to perform Bayesian inference on such models automatically. However, due to the expressiveness of PPLs, it is very challenging to design general purpose inference algorithms for models specified in Turing-complete PPLs. Applications tend to build their own custom inference machines, which is time consuming and labour intensive. This proposal puts forward a recently introduced framework—the involutive Markov chain Monte Carlo (iMCMC), as a candidate for inference in a PPL, namely Turing. Such a framework would enable the automation of an implementation of a correct inference algorithm from user specifications in Turing. Moreover, Turing readily supports differentiable programming (in their HMC implementation) which is the main technical requirement of iMCMC. Early iMCMC developments in Gen show promising results compared to standard black box styled inferences. Hence, it would be desirable to have an implementation in Turing.</p>

<p>A leap forward in the accuracy of forecasting problems in chaotic time series has been recently obtained using Echo State Networks (ESN). This is a novel kind of recurrent neural network with a randomly connected hidden layer, called the reservoir, and an adaptable output layer trained using a simple regression algorithm. This architecture has the advantage of faster computation and presents less parameter tuning then a standard neural network. Since their inception there has been a lot of effort towards the improvement of the model and applications have been found in several fields of study: anomaly detections in geophysics , short term predictions in astrophysics and reconstruction of chaotic attractors among others.</p> <p>This project aims to build a comprehensive Julia package in which are implemented the majority of the variations of the ESNs presented in the literature. Such an effort would represent the first attempt to create a library for this kind of models in any programming language, and would reap benefits not only for the consequent simplicity of application, but also for showing the actual state of the art in a family of models that has yet to go mainstream.</p>

<p>The purpose of the project is the implementation of an online Greek mail dictation system. In practice, the user will dictate the mail that wants to send and the speech will be converted into text. The system's performance will be improved through the training of a personalized acoustic and language model. Extra features will be supported, such as special dictation commands and replay of the final email for verification.</p>

This projects aims to bring the well known ColBERTv2.0 information retrieval system to Julia! In this project, we aim to implement various components of the ColBERT model, with the major focus on indexing and search. Being a late-interaction retrieval system, ColBERT isolates all the computations between the queries and documents to enable pre-computing the document embeddings offline. Combined with a RAG pipeline as an embedding and retrieval backend, ColBERT can make the usage of existing AI tools really powerful. As part of the project, we also aim to apply this model to Julia's own extensive documentation (over thousands of packages), and integrate ColBERT with AIHelpMe.jl and PromptingTools.jl. Combine this with a locally hosted LLM, all Julia users get a powerful, modular, and easily hackable local AI assistant, allowing them to integrate ColBERT with their own RAG pipelines. The main deliverable of the project will be a new package (called ColBERT.jl), with the implementation design inspired from the original Python implementation of ColBERTv2.0 (i.e the centroids + residuals compression method). The project will also have an extensive number of addons, including, but not limited to, implementing the training module for this package, which will allow users to train their own ColBERT models instead of relying on pretrained checkpoints, and distributed training (using existing art like DaggerFlux.jl).

<p>Problems with memory management and array access by unchecked index or pointer introduce significant risk. Especially if it comes to the OS kernel. LDV project aims to improve Linux kernel and I want to correct issues with its environment model for kernel.</p> <p>As a results of my work I expect to find and fix a number of bug in one of the last kernel releases and improve LDV verification framework.</p>

I aim to upgrade Visdom from a single-user visualization tool into a scalable, multi-user system. Visdom currently operates on a shared global state and lacks features like user context, authentication, workspace isolation. I plan to implement a JWT-based auth system and refactor the architecture to support user-scoped state. Deliverables are multi-user system featuring isolated environments, role based access control, basic collaboration and presentation features, controlled experiment sharing.

<p>The bhyve hypervisor does not have any sound card emulation at the moment. This project is proposed by Peter Grehan and aims to implement the High Definition Audio Specification which is one of the best performant and supported device in the nowadays operating systems. Some of the reasons to have this device are: support for multiple input and output streams, up to 16 channels per stream, command-response codec communication, DMA channel for each stream and it is interfaced with the PCI bus. The main goal is to have a functional implementation by the end of the summer that will allow playing sounds inside a bhyve virtual machine.</p>

Running Kubernetes should be done with the utmost security in mind. KubeArmor provides a policy based system to increase the quality of security in a Kubernetes cluster. Users need to define their own policy when implementing KubeArmor. Most users might find it challenging to define the most appropriate and effective security policy based on their use case. Currently there are limited ways for users to actually get observability data that can be used to assist them and provide better reasoning when creating an optimal security policy. The solution to this problem is by providing visibility telemetry events to show pod or container observability data. Observability data can consist of processes executions, file system accesses, and network accesses. This information is expected to assist users to create an optimal security policy. This solution is implemented by developing and deploying a Kubernetes service that will connect to the KubeArmor relay and get the events from the daemonsets. These events will then be aggregated and stored in the database for future use. Users can get the data by using the KubeArmor CLI program which will be extended to accommodate this solution.

<p>QuTiP is best known for solving open quantum system dynamics. At the same time, it also has a Quantum Information Processing (QIP) submodule representing ideal quantum circuits. A tempting ideal to combine them is introducing random noise into the circuit by linking the circuit back to the Hamiltonian driving the evolution of the qubits. This could then be used to study the possible noise occurring in the experiments and how it will influence the result.</p>

<p>With booming amount of information being continuously added to the internet, organising the facts and serving this information to the users becomes a very difficult task. Currently DBpedia hosts billions of such data points and corresponding relations in the RDF format. RDF is a directed, labeled graph data format for representing information in the Web. SPARQL is a query language for RDF.</p> <p>Extracting data requires a query to be made in SPARQL and the response to the query is a link that contains the information pertaining to the answer or the answer itself. Accessing such data is difficult for a lay user, who does not know how to write a query. This project will try to make this humongous linked data available to a larger user base in their natural languages(now restricted to English). The primary objective of the project being to be able to translate any natural language(English) question to valid SPARQL query.</p>

<p>Building a library that would take customer and transactional data as input to identify Good/Bad Loans, Frauds and other risk management activities using Machine Learning Algorithms</p>

The project aims to bring the Prometheus Agent's DaemonSet mode in Prometheus Operator to full production readiness. This involves addressing missing functionalities, improving configuration validation, enhancing reliability, and providing comprehensive documentation. The ultimate goal is to offer a scalable, efficient, and fully supported way to run Prometheus Agent across Kubernetes nodes. This is achieved by integrating remaining features like using CEL for CRD validation to provide immediate feedback along with adding runtime checks as a fallback for older Kubernetes versions. Furthermore, Investigating and potentially implementing EndpointSlice for scalability and extending integration tests to multi-node Kind clusters. Finally, adding detailed documentation and example manifests.

<p>The availability of depth sensors such as 3D laser range finders or the Microsoft Kinect provide dense 3D measurements that typically consist of millions of points. In robotics applications like object manipulation or teleoperation, it is often crucial to interpret this massive amount of data in real-time. The ability to segment the rgbd scene captured in a point cloud file into planes, spheres and cylinders, from within OpenCV will undoubtedly be a useful utility for scene understanding.</p>

<p>Likelihood-based methods have many appealing features, but they are difficult to use for models which do not have a closed-form, tractable likelihood function. When these models can be solved numerically and simulated, indirect inference can be used. In a Bayesian context, this approach is called Approximate Bayesian Computation (ABC). ABC methods using an extension of the Metropolis-Hastings algorithm (Marjoram et al 2003) allow sampling from the posterior using simulation, but are not effective for larger dimensions. Hamiltonian Monte Carlo (Neal 2012) and related algorithms can overcome these problems when the likelihood is available. Hamiltonian Approximate Bayesian Computation combines the ABC methodology with the Hamiltonian approach.</p>

The proposal aims to integrate the Julia package Agents.jl (for agent-based modeling) with ReinforcementLearning.jl or Crux.jl, enabling intelligent agents in simulations to learn optimal policies via reinforcement learning. The integration will allow for dynamic, reward-based behaviors rather than static rule-based actions. The project involves updating the chosen RL library for compatibility with the latest Julia version, building a lightweight interface (ABMEnv) to connect it with Agents.jl, and showcasing the integration through example applications like herding, civil violence, and predator-prey models. This work will expand the capabilities of Julia’s ecosystem for researchers in fields such as economics, social sciences, and robotics, while offering a standardized and accessible way to combine RL and ABM.

Graph Neural Networks (GNNs) are powerful tools for processing graph-structured data, with applications spanning social networks, recommendation systems, molecular modeling, and beyond. The Julia package GraphNeuralNetworks.jl, based on Flux.jl, provides access to common graph convolutional layers, offering a flexible framework for building GNN models. Sparse computations are essential for the efficient implementation of GNNs, since many operations, such as propagating node features, can be expressed as the multiplication of a sparse adjacency matrix with a dense feature matrix. Improving these operations through enhanced GPU kernels and fused operations is critical to reducing memory overhead and increasing the overall performance. This project aims to address these limitations by enhancing CUDA.jl’s sparse matrix support and integrating these improvements into GraphNeuralNetworks.jl, ultimately enabling faster and more scalable GNN computations in Julia.

<p>Density functional theory (DFT) is a key driving force in modern quantum chemistry, with applications ranging from materials science to drug discovery. From a numerics perspective it amounts to solving a coupled system of non-linear partial differential equations. This is the setting of the density functional toolkit DFTK.jl, a Julia package for solving DFT problems.</p> <p>Can we make a full DFT solver differentiable via automatic differentiation (AD)? There are at least three reasons why this is an interesting problem:</p> <ol> <li>It will enable faster model design iterations for chemists, providing useful model derivatives automatically.</li> <li>It will enable research into novel combinations of machine learning and DFT.</li> <li>It is technically challenging from an AD perspective. We expect to raise rough edges and collaborate on smoothing them out. We believe it has the potential to become a showcase of the Julia ecosystem and AD compositionality on a large scale.</li> </ol>

The main points of my proposal are as follows: 1. Add more and more datasets: To enrich MLDatasets.jl with relevant Graph Datasets. 2. Build a tight integration with GraphNeuralNetworks.jl The graph datasets are built to be used with libraries like GraphNeuralNetworks.jl(still in it’s beta stage), we can build the MLDatasets.jl graph datasets to be first class compatible(usable without little to alterations). 3. Consistent Structure The data for MLDatasets are sourced from different websites, who in turn use their own variable names and specifications. Currently MLDatasets does-not enforce any fixed api for data-access, so an end-user needs to refer to the docs for each dataset. Maintaining a fixed API for a dataset will make the repository more usable. 4. Loading of Large models: Most of the modern day datasets does not fit into [personal] computer memory. The dataset needs to chunked before it can be loaded into memory. MLDatasets.jl currently lacks methods for chunking, so developing methods for loading large [Graph] Models in memory necessary.