Find open source projects actively accepting contributors. Search repositories, filter by program milestones, difficulty tags, or tech stack.
Use our Orbit AI Matcher to find out! Get instant matching scores based on your developer skills, preferred frameworks, and contribution experience.
Convert your selected open-source project into a winning GSoC, LFX, or Outreachy application using Proposal Studio.
This project aims to extend the classbound R package into a general framework for exploring and comparing classification boundaries across a wide range of models and datasets. The project will develop a modular R package that supports additional classification algorithms, including user-defined methods, along with flexible data simulation utilities and support for real datasets. It will also extend functionality to high-dimensional settings using tour-based visualization techniques, allowing users to interactively explore data structure and corresponding decision boundaries. The package will be designed for extensibility and will include comprehensive documentation, examples, and a vignette. Final deliverables will include implementations of additional classifiers, enhanced simulation capabilities, high-dimensional visualization support, and complete documentation, enabling users to better understand and compare classification models through their decision boundaries.
<p>A GUI is very essential for TARDIS so as to allow researchers to easily investigate the important information about a simulated supernova model, without knowing its exact inner data structure. But the existing TARDIS GUI is outdated and is implemented in Qt which is no longer an ideal framework. These days, since most of the analysis work is done in Jupyter Notebooks, thus TARDIS GUI can also be brought to the notebooks by using Jupyter widgets and interactive plotting libraries.</p> <p>Moreover, the existing GUI is limited in its functionality, providing only shell & line information. There is lot of other information that can be analyzed from a simulation model which will be made accessible under this project, by means of new widgets. Besides, this project also aims to make it easier for users to try out these widgets themselves. For this, several example notebooks will be embedded in docs, <em>thereby showcasing the unprecedented possibilities with TARDIS!</em></p>
<p>LLVM Project is a collection of modular and reusable compiler and toolchain technologies. It supports various high-level languages and hardware architectures. LLVM IR is an SSA-based intermediate representation whose design helps in the easy and efficient implementation of various compiler optimizations and transformations. Alive2 is a tool that checks whether the transformation of an instruction set into another instruction set is semantically correct or not. It is a verification framework for LLVM IR, which checks whether the source and target IR are equivalent or disproves by providing a counterexample. Alive2 can reason about a wide variety of intraprocedural optimization problems, including control flow, memory, vectors, and floating-point. Alive2 has detected some fundamental issues in the LLVM-IR, which lead to mis-compilation and unsound optimizations. In this project, I propose to solve some of these issues in the LLVM IR by making some changes to the LLVM compiler infrastructure code.</p>
Kolibri's QTI 3.0 viewer currently renders only Multiple Choice and Text Entry, leaving the majority of the specification's interaction types unsupported. This project extends the plugin to implement six additional types: Inline Choice, Match, Order, True/False, Gap Match, and Fill-in-the-Blank — each requiring careful integration with the existing SafeHTML rendering registry and QTIVariable response system. The core technical challenges are accessibility and interaction fidelity: drag-and-drop interactions (Order and Gap Match) require a fully keyboard-native select-to-place model as the primary interaction, with SortableJS-based drag as the pointer enhancement. Match Interaction requires a directed-pair association model rendered as an accessible grid. All six types must support RTL languages, touch devices, and low-end hardware — not as retrofits, but as first-class constraints from the start.
<p>In this project we intend to integrate publicly available -omic and clinical datasets using natural language processing techniques. Combining genomics data with physiologic read-outs may be effective in creating robust machine learning and data analysis pipeline. The example microarray gene expression data can be downloaded from GEO (<a href="https://www.synapse.org/#!Synapse:syn5612563" target="_blank">https://www.synapse.org/#!Synapse:syn5612563</a>) and physiologic data from eICU (<a href="https://eicu-crd.mit.edu/" target="_blank">https://eicu-crd.mit.edu/</a>). The idea is to map phenotypic terms to causal genes (for sepsis) and follow the SIRS timeline to form the integrated data set. After that robust machine learning models can be formed using the integrated data and compared with already existing models.</p>
<p>The Mesh-RCNN technique was recently introduced to estimate 3D models from natural 2D images. It uses a combination of an object detection module which works on real-world images and produces a triangle mesh giving the full 3D shape of each detected object. Methodically, Mesh R-CNN comprises different modules for mesh prediction which provide functionality that would be cool to have in TensorFlow Graphics. TensorFlow Graphics already contains implementations of some key building blocks of Mesh R-CNN, like graph convolutions and the chamfer distance by the time of this writing. However, some other crucial parts of the Mesh R-CNN approach need to be implemented als modules into TensorFlow Graphics. This projects provides the means to implement Mesh R-CNN in TensorFlow Graphics and train it using an TensorFlow Dataset implementation of Pix3D, a Dataset for Single-Image 3D Shape Modeling.</p>
Develop aeon-neuro to provide structured tools for machine learning from neural data. This project will focus on implementing algorithms for EEG classification by building on the multivariate classification algorithms outlined in Rushbrooke 2023. This paper demonstrates that existing time series models implemented in aeon can successfully classify patients from healthy individuals using frequency domain features alone, eliminating the need for detailed time domain feature selection. In addition to applying existing machine learning models to EEG datasets, we will further develop aeon-neuro to be more accessible to the scientific research community by interfacing it with existing data formatting standards (BIDs) and EEG analysis libraries (MNE). Alongside these primary outcomes, we will adhere to best practices in research software development, including writing well-test code, consistent documentation, and user-facing examples/notebooks.
This project aims to improve Jenkins CI/CD troubleshooting by creating a domain-specific Large Language Model (LLM). The core task involves fine-tuning an open-source LLM (e.g., Phi-4/Qwen 2.5) on structured data derived from ci.jenkins.io build, logs and metadata, using resource-efficient techniques like QLoRA with Unsloth. To ensure feasibility despite potential API limitations, a primary strategy involves heuristic/manual data curation, supplemented by LLM assistance where possible. A baseline Retrieval-Augmented Generation (RAG) system incorporating Jenkins documentation will be implemented to enhance factual accuracy. Key deliverables include the fine-tuned model (GGUF for Ollama), a basic FastAPI backend, the RAG pipeline, and an evaluation report incorporating both automated metrics and qualitative mentor feedback
Using machine learning, this project will develop an image processing pipeline that takes a user submitted image from Boston’s 311 app and helps the resident submit a corresponding non-emergency city request. By automating this process, this service can be accessible and more intuitive for a wider range of audiences, who may not have English as their first language or be well-versed with using technology, while also consuming less of the city’s resources due to erroneous submissions. Curating a dataset from the previously submitted 311 system photos, I will prepare a large image collection that will then be used to develop a CNN for image classification with the TensorFlow library. Then, I will develop a large language model trained on past service requests that can help draft a service request for the user that is customized to the specific situation at hand. These models will be saved, compressed, then deployed to an accessible API endpoint.
<p>Combustion has enabled our technologies in electricity production, heating, transportation, and industry. Advancements in combustion technologies depend on computer simulation software; one such software is Cantera, an open-source chemical kinetics program. This project proposes the development of a new solver for Cantera to equip users with a tool to simulate and analyze 0-D steady-state combustion systems. These types of systems are idealized theoretical models that are used as good approximations of certain real combustion systems, or as building blocks towards more complex ones. The development of this solver will begin with mathematical modeling of the physical laws that define 0-D steady state combustion systems. It will require the application of numerical analysis techniques to quickly approximate accurate solutions. Then, it will be coded into the Cantera source code and thoroughly tested before release as a new feature in the next official version of Cantera.</p>
As Zulip plans to refine the settings system using the User Groups based permission model, there is a requirement to update the settings interface for User Group settings to align with the current plan, and the addition of new settings related to user groups is more straightforward. In this project, one of my goals would be to enhance and improve the complete User Group settings overlay so that it looks more like stream settings so that working with settings on user groups becomes smoother. Apart from the above migration, another goal would be to actively contribute towards the migrations of settings systems to the user group-based permission model by actively collaborating on the major issues regarding this. This would require working on a lot of areas description of which is included in relevant issue threads and the proposal. Having contributed majorly towards settings-related areas, I'll also try to contribute to areas like message viewing, markdown, etc., and will try to resolve critical issues related to those.
WasmEdge provides several AI frameworks as WASI-NN plugins to enable the power of AI/LLM applications for developers and users. We are always eager to add new backends to improve coverage of all models and hardware. BitNet.cpp, released by Microsoft, offers the ability to run 1-bit LLMs quickly without a GPU. We would like to support this framework so that people with limited resources, such as CPU-only hardware, can enjoy the amazing world brought by LLMs. Expected Outcome: i. A new WASI-NN plugin supports BitNet. ii. Use the pure C++ interface from BitNet without any Python dependencies. iii. The plugin must run the model listed in the BitNet repository, e.g., BitNet b1.58 2B4T - Scaling Native 1-bit LLM. iv. A tutorial and example for demonstration. v. A CI workflow for building, testing, and releasing the built assets.
This study aims to show the capabilities of QML especially QCNN for classifying the HEP image datasets. QCNN can be completely quantum or can be a hybrid with classical. The aim is to implement both. We will use quantum variational classification instead of the final FC classical layers in the quantum setting. This will give more depth about the quantum power that can be used in the near-term future. I summarize the required deliverables that will be achieved before the official end. Implementation of some chosen classical ML models as a baseline using TensorFlow. Implementing QCNN for fully quantum and hybrid architectures including different data encoding schemes (angle, amplitude encoding etc.). Jupyter notebooks will be used as a testing environment but the final python modules will have a .py extension. These notebooks will then be extended as a form of documentation for beginner contributors. Comparison of classical ML baseline models with QCNN and documenting the results.
<p>The existing query builder in InterMine requires some experience to obtain statistical data about the different mine categories, which can become overwhelming for new users. For instance, for a user interested on searching data in HumanMine using it’s query builder, he or she would need to browse through the different classes and attributes of the ontology, choosing between the available fields and adding different constraints over each of them in order to get the desired output. This requires the user to have a decent knowledge of the model schema in order to successfully build a correct query giving the expected query results. For new users, this workflow can become, indeed, overwhelming to search for specific information in the data.</p> <p>For this reason, the goal of this project, is to implement a faceted search tool to display the data from the InterMine database, allowing the users to search easily within the different mines available around it. without the requirement of having an extensive knowledge of the data model.</p>
<p>JdeRobot's DetectionSuite is primarily used for fair comparision of different Deep learning models on any framework or dataset preferred. This project proposal aims to expand DetectionSuite’s current functionality to support more deep learning Frameworks such as Tensorflow, keras and Caffe and more datasets such as Imagenet, COCO, etc. It also aims to add Segmentation support to this tool along with various evaluation metrics in order to compare different detection models. For easy installation, this project will also improve current documentation and create debian packages.<br> <a href="https://drive.google.com/open?id=1HSByDVfi6ZOqKVTAJTiUb11ATFTalp-P" target="_blank">Link to Proposal</a><br> <a href="https://jderobot.org/GSoC-2018#Project_.233:_Improving_DetectionSuite_deep_learning_tool_including_segmentation_and_classification_tools" target="_blank">Project Link</a><br> <a href="https://github.com/JdeRobot/dl-DetectionSuite" target="_blank">Github Repo</a></p>
<p><em>Interoperability Model</em>Port the gen-api-models tool to OpenAPI v3</p> <p>I spoke with Mentor/managers on Slack (Leonardo Favario and Roberto Polli) Which I thank for giving me all the answers concerning the project. They told me that they must be maintained: I update Kong(Open Source API Gateway) Rate Limiting and edit it to fit the current system, I add a lot of test code(unit test, etc..) and update documentation. API rest maintenance and adapt them to the latest versions of OpenAPI v3. I'm in contact with mentors and managers, I hear them at least 3/4 times a week. Every time they update me on something about the project, I report it</p> <p><a href="https://docs.google.com/document/d/1Navvdxo6QONWg6zz7LaXka4iDZt6ERTIaisFCgiQbQA/edit?usp=sharing" target="_blank">https://docs.google.com/document/d/1Navvdxo6QONWg6zz7LaXka4iDZt6ERTIaisFCgiQbQA/edit?usp=sharing</a></p>
<p>Sequential Monte Carlo (SMC) methods are powerful alternatives to standard Markov chain Monte Carlo (MCMC) for sampling from the posterior of complex static Bayesian models. SMC is naturally adaptive, parallelisable and capable of dealing with complex and multimodal targets. However, SMC is generally seen as more difficult to implement than MCMC. R packages like RcppSMC could help to make these methods more accessible.</p> <p>An estimate of the marginal likelihood (normalising constant or evidence) can be obtained as a by-product of SMC (see e.g. Del Moral et al (2006)). Adjusting the current RcppSMC package to allow for estimation of the marginal likelihood would be useful for model choice and for exact-approximate methods such as particle MCMC (Andrieu et al., 2010).</p> <p>There is a plethora of choices for advanced SMC implementations since this is an active area of research. Some of the possibilities have been described in more detail in the timeline below.</p>
<p>The main goal of the project is to provide the user of DeepPavlov Framework with an out-of-box solution for relation extraction. I would consider relation extraction as a multi-label classification task and design a pipeline that could reuse the existing DeepPavlov components (such as NER with incorporated tokenizer, already implemented MT-BERT, and others) as much as possible. I would train several relation extraction models with different parameters (for example, different training data, classifier, classifier inputs, inputs encoding, number of relations to be extracted, etc), add them to the DeepPavlov storage, and give the user an opportunity to load them (and, occasionally, additionally train with his/her own training data) and find the one that would suit his/her data the best way. The user's input would be a text and a config with RE parameters, and relational triples would be the output. Thus, the whole preprocessing, data encoding and model application would happen inside the module, while the overall relation extraction component remains clear and convenient for the user.</p>
<p>Godot's high-level multiplayer approach makes it uniquely easy to create client-host and true peer-to-peer (P2P) models of multiplayer gaming. However, outside of local area networks Godot struggles to actually create these multiplayer models because it lacks any sort of NAT traversal. I plan to implement WebRTC. The WebRTC standard will be able to provide support to Godot's existing multiplayer in making these sorts of peer-to-peer games.</p> <p>It will also make multiplayer viable in the HTML5 target platform, and allow for easy interfacing with non-Godot peers that use WebRTC. Additionally, the support I build for WebRTC will make it easy to later add WebRTC MediaStreams, allowing for in-game voice chat. (However, building MediaStream is less-often needed for games so is considered to be low priority and non-essential in my project. See project proposal for details.)</p>
Supervised learning might be challenging in cases where there are extremely few known instances in a given category. This is a common occurrence in the case of Strong gravitational lensing images when samples from one or more classes are scarcer than others. On the other hand, when there are few data labels available for supervision, Self-Supervised Learning (SSL) has been shown to perform better than conventional supervised learning. SSL can also create meaningful representations from very vast unlabelled datasets that would be challenging or impossible to label manually. Additionally, transformer-based architectures recently have emerged as a paradigm to solve a wide range of tasks in computer vision from image classification to object detection. They have a number of advantages over traditional methods, including dynamic weight adjustment based on input features, global context, and improved generalization. Thus, the project aims to train transformer architecture in Self-Supervised Learning fashion to learn generic features. These features could then be used in different downstream tasks such as image classification or regression.
Clang, part of LLVM, supports multiple languages and functions as a library, fostering a compiler-assisted ecosystem. Its accessible codebase and LLVM's JIT advancements enable innovative C++ processing methods, though challenges like incremental compilation and compile/link-time optimizations persist. Incremental compilation processes code in segments, building translation units and enabling efficient interpreters. Clang-Repl, using Orcv2 JIT within the same process, exemplifies this concept but faces limitations on resource-constrained devices and risks process-wide crashes on user code failure. This project aims to transition Clang-Repl from an in-process to an out-of-process execution model to address: 1. **Resource Constraints**: The current model requires substantial resources, limiting use on devices like Arduino Due. 2. **Crash Resilience**: In-process crashes can bring down the entire Clang-Repl process, affecting stability and user experience.
Modern AI systems relying on Retrieval-Augmented Generation (RAG) often suffer from stale data due to batch-based updates. This project, PyDebeziumAI, aims to solve this by enabling real-time synchronization between databases and LLM pipelines using Debezium Change Data Capture (CDC). The project will extend pydbzengine to stream CDC events directly into LangChain and LangGraph by transforming database changes into semantic Document objects and updating vector stores in real time. Key deliverables include: - A Python library integrating Debezium with LangChain/LangGraph - Real-time vector store synchronization (Chroma, PGVector, Milvus) - Pluggable document transformation and ID strategies - End-to-end examples (RAG chatbot, reactive agent, dashboard) - Full documentation, testing, and PyPI release This will provide a production-ready bridge between CDC systems and AI frameworks, ensuring always up-to-date LLM context.
DigiKam’s Advanced Search is powerful, but translating a natural-language request into the right combination of structured filters can be difficult for many users. This project will add a natural-language input to digiKam’s Advanced Search so users can describe queries such as “landscape photos with red labels taken in Paris last summer,” which will then be converted into editable search criteria such as tags, labels, dates, and locations. The implementation will use a lightweight local language model, structured intent parsing, capability-based resolution into supported digiKam search fields, ambiguity handling for unclear terms, and caching for repeated queries, while reusing digiKam’s existing model-management infrastructure. The deliverables include the Advanced Search UI integration, the local inference and parsing pipeline, criteria resolution and validation, safe clarification behavior, and end-to-end integration with digiKam’s existing Advanced Search workflow.
This project comprises two parts: prototyping a real-time Python pipeline for object detection on point cloud data, and integrating it into LivePose. The goal is to build a functional pipeline that reads streaming point cloud data from various sources (e.g., Orbbec Femto Megas, Intel RealSense, stereo cameras), runs deep learning inference using models like PointNet++, applies filters for smoothing and tracking, and streams the results via OSC. During prototyping, the system will be evaluated across multiple models, input types, and hardware configurations to assess real-time performance and reliability. Once validated, the pipeline will be integrated into LivePose, leveraging its modular infrastructure for input handling, filtering, and output. By the end of the project, LivePose will support end-to-end processing of streaming 3D point cloud data in at least one working configuration, with clear documentation and performance benchmarks to guide future development and extension.