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.
One of the goals of this project is to utilize Deep Learning models to better interpret underwater sounds. We are required to train neural networks streams of audio data to separate marine mammal vocalization sounds (orcas and humpback whales) from other background noise like that of ships and other vessels. Next, we have to train two different neural networks- one for classifying the whale vocalization sounds and the other one to classify ship noises. The second part of the project involves building an efficient vessel monitoring system. We will use AIS data to detect and track larger commercial vessel and some private yachts. The non-AIS boats can be tracked using RADAR and camera data using Marine Monitoring system. Once we are done training all the models, we can deploy the detectors on the cloud and on the Jetson Nano.
EchoGem introduces a novel batching engine designed to answer multiple questions about the same source parallelly to reduce response times heavily. It does so by building a modular smart batching engine for Gemini that teaches the model to think in context-aware batches. What sets EchoGem apart is its focus on modularity and testing—each component is designed to be independently swappable and improvable allowing for open testing and matching different strategies for different parts of the engine. Instead of a naive sliding window or top-k search approach, EchoGem uses semantic clustering of both chunks and questions to form intelligent batch groups. It is also committed to measurable, reproducible gains. Every design decision—whether it's a new chunking strategy or a different context ranking model—can be rigorously backtested using a structured evaluation suite.
<p>The PPI (Poverty Probability Index) is a powerful tool to measure poverty index and also a yard stick to measure the progress and impact of poverty outreach programs. To reduce the amount of human intervention in surveying for more accurate and quicker results, I have submitted a proposal in which I plan to make an Android application which will use Machine Learning, Computer Vision through the Google Cloud Vision API and Tensorflow library to perform image analysis and aid the surveying process. The app will render the relevant PPI using the Apache-Fineract SPM API. Further, images may be uploaded to the Cloud Vision Models or fed to the embedded Tensorflow models which are made available using Firebase ML Kit to obtain a report of various objects detected and inferences derived relevant to PPI questionnaire.</p>
Sampling and volume computation of convex polytopes is a challenging computational problem with many applications in inference from linear constraints [1]. Thus, sampling algorithms are essential in many scientific fields. Biology in general, and most particularly metabolic modeling, is among those scientific fields that need such approaches the most to address their modern challenges. Flux Balance Analysis (FBA) is a typical example of a constraint-based metabolic model (CBMM) [2]. However, flux sampling is considered a superior methodology. Python package dingo provides the Multiphase Monte Carlo Sampling (MMC) algorithm to sample from the flux space of a metabolic network. However, the preprocessing methods implemented in dingo need major improvements regarding memory allocation. The goal of this project is to improve the preprocessing methods in dingo to support faster sampling and rounding pipelines.
This project will build a portable Kafka Streams runner for Apache Beam so that Beam pipelines can run on Kafka Streams using the Beam Fn API. The runner will translate portable Beam pipeline protos into Kafka Streams topologies and support a working skeleton for core transforms: Read, stateless ParDo, GroupByKey, Combine, non-merging windowing, and Flatten. The implementation will focus on correct watermark propagation, commit-based bundle boundaries, exactly-once semantics, and integration with the existing Beam portability stack. Deliverables include a new runners/kafka-streams module, runner entry points, a pipeline translator with transform-specific translators, watermark and bundle management, tests using TopologyTestDriver and real Kafka-based integration coverage, and documentation for the Beam community.
The way humans interact with each other occurs in multimodality. We not only articulate words, but also we show them. Expressing different concepts such as time, place, and emotion comes with speech and some movements called body gestures. In this study, we propose a multi-modal method that captures different patterns of body gestures aligned with an articulated time expression such as "from beginning to end". Our proposed architecture lies on two neural networks, the Compact Transformer and the Long-Short Term Memory (LSTM). Compact Transformer which is the current State-of-the-art structure for temporally distributed images also performs well over low-resourced data. LSTM performs well over temporal data with long-term dependencies. The hypothesis of this project is the existence of a relation between the body gestures and the time expressions we speak and using neural networks we try to accept or reject this hypothesis.
<p>Linear Programming (LP) based sparse learning methods, such as the Dantzig selector (for linear regression), sparse quantile regression, sparse support vector machines, have been widely used in machine learning for high dimensional data analysis. Despite of their popularity, their software implementations in R are quite limited. Our GSOC project aims to develop a new R package -- PaRametric sImplex Method for spArse Learning (PRIMAL), for the aforementioned LP-based sparse learning methods with the following two key features: 1) It provides a highly efficient optimization engine based on the parametric simplex method, which can efficiently solve large scale sparse learning problems; 2) Besides the estimation procedures, it provides additional functional modules such as data-dependent model selection and model visualization.</p>
This project aims to develop a comprehensive Hamiltonian Library for QuTiP (Quantum Toolbox in Python). Currently, QuTiP users must frequently recreate common quantum system Hamiltonians from scratch, leading to redundant work and inconsistencies. The library will provide ready-to-use implementations of frequently used quantum systems like Jaynes-Cummings, Rabi, Dicke, and various spin-chain models. Key features include a standardised API, JSON serialisation for cross-platform compatibility, time-dependent parameter support, and LaTeX representation capabilities. The deliverables include a core Hamiltonian base class, at least 10 implemented quantum models, comprehensive documentation, tutorials, and full integration with QuTiP's existing functionality. This project will significantly enhance user experience and reduce redundancy in quantum simulations.
<p>360° video is getting popular during the past years and it is predicted to evolve [1]. Compared to virtual and augmented reality, 360° video is much more accessible for viewers. The cameras are getting cheaper, and many video sharing services have an option for 360° video support (YouTube, Vimeo, Facebook) as well as applications (e.g. VLC).</p> <p>360 video requires special transformations to be viewed: the most popular way to store the data at ERP and CMP projections.</p> <p>The goal of the project is to implement and optimize ERP and CMP conversion algorithms using GLES. The algorithm will perform 360° video transformations between ERP and CMP. [1] <a href="https://images.app.goo.gl/iVyk3qkdxmR7CNGA8" target="_blank">https://images.app.goo.gl/iVyk3qkdxmR7CNGA8</a></p>
<p>Current limitations with the variant detection using wbtools (and entity extraction in Wormbases’s AFP pipeline) is that it relies on regular expression and string matching with database. This approach can not be scaled to the entities or variants which do not have a clear pattern or are novel - especially relevant when there’s a lack of common consensus in SNP nomenclature.</p> <p>This proposal, in short, aims to increase entity detection capabilities of the AFP pipeline (and variant or strain detection capabilities of wbtools) by integrating more data models in the current pipeline, and for cases which aren’t possible to be generalized using a data model, use the present curated databases to train Named Entity Recognition (NER) using BERT to extract data from papers. Then, validate the discovered SNP by using NCBI’s Basic Local Alignment Search Tool.</p>
<p>ArviZ is a Python package for exploratory analysis of Bayesian models, intending to provide backend-agnostic tools for diagnostics and visualizations of Bayesian inference in Python. Generally, Bayesian inference generates numerous datasets that represent different aspects of the model. To simplify the handling, referencing, and serialization of data generated during Bayesian analysis, ArviZ uses a couple of data structures: <em>xarray.Dataset</em>, <em>arviz.InferenceData</em> and <em>netCDF</em>.</p> <p>InferenceData objects are central to ArviZ and most ArviZ functions take InferenceData as input. However, its functionality is still quite limited. The main aim of the project is to extend some methods from <em>xarray.Dataset</em> and to create specific InferenceData methods. Special attention would be given to test and document these new functionalities with examples.</p>
Tardis is an open-source project designed for rapid spectral modelling of supernovae. Astrophysicists widely use it to model supernovas. When new astrophysicists start using the Tardis, they will use the help of past projects to understand the simulated spectra. So it is crucial to continuously update the simulated spectra of past projects as the code of Tardis gets modified. Currently, three past projects are stored in the tardis-setups repository. Each time there is a change in the codebase of Tardis, it is hard to run the projects and update the spectra manually. So a CI/CD framework should be created to automate comparing and publishing the results. To achieve this, we intend to use GitHub Actions. The existing projects have different ways to run the simulations. So the existing projects will be modified to be integrated into the pipeline. Guidelines will also be put in place to allow future contributors to add more projects.
<p>A large number of interacting bodies is the key component in explanation of many physical phenomena. From the gravitational interaction of stars and planets to molecular dynamics and electron gas in quantum mechanics, the N-body problem formalism plays an essential role in modeling those systems. The goal of this project is to create tools for conducting N-body simulations in a fast, accurate and reliable manner. Often the interaction of bodies can be described by means of differential equations (DEqs) so that the task of finding the state of the system at a particular time leads to solving those equations. The packages of JuliaDIffEq organization were developed exactly to supply efficient Julia implementations of solvers for various DEqs. The current project describes an extension of DiffEqPhysics.jl package, which will allow researchers to solve the N-body problems in different physical models and applications.</p>
This project focuses on transforming Mogan STEM into a web-based, multi-user collaborative platform. Currently, the lack of a functional WebAssembly (WASM) version and real-time synchronization limits its use in modern collaborative research. To solve this, I will compile Mogan STEM using Emscripten, implementing a custom virtual file system to ensure stability. Furthermore, I will develop a real-time collaboration core based on Operational Transformation (OT) and WebSockets. This architecture allows multiple users to edit complex scientific documents simultaneously with "local-first" responsiveness and guaranteed global consistency. Deliverables include: 1. A stable browser-based version of Mogan STEM. 2. An OT-based synchronization server. 3. A complete end-to-end prototype supporting real-time multi-user collaborative editing.
This project investigates the fusion of classical data encoding onto quantum models using contrastive learning techniques. The objective is to develop a function for encoding classical data onto quantum feature spaces, enabling their representation in a format compatible with quantum processing. By contrasting positive (similar) and negative (dissimilar) pairs in the feature space, contrastive learning facilitates representation learning. The trained embeddings will be assessed against standard encoding approaches using quantum machine learning (QML) models, providing insights into their performance. Through this exploration, the aim is to evaluate the efficacy of quantum embeddings for classical data and compare them with traditional encoding techniques. By harnessing the capabilities of quantum computing and contrastive learning, this endeavor contributes to advancing classical data encoding on quantum platforms, potentially unlocking novel applications in quantum-enhanced machine learning.
This project aims to implement comprehensive support for correlated subqueries in Apache DataFusion by applying Hyper's 'Unnesting Arbitrary Queries' framework. With a complete solution that can transform arbitrarily nested correlated subqueries into standard join operations through a systematic approach. The implementation will include enhancing the binding phase to handle multi-level nesting, implementing the core dependent join transformation algorithm, applying operator-specific pushdown rules for all relevant operators, and optimizing redundant operations. This approach provides a general solution that works for all types of subqueries (scalar, EXISTS, IN) and handles complex cases involving aggregations and multiple levels of correlation. By adopting this proven theoretical framework, DataFusion will gain robust SQL compatibility and significantly improved analytical capabilities, enabling users to express complex data relationships efficiently.
<p>Property Wrappers are responsible for wrapping common getting and setting boilerplate and also for storing any auxiliary helpers. Often, these helpers are the same across different instances of the wrapper. Thus, having to store it in each individual wrapper instance is a waste that should be avoided. We can solve this by introducing a static storage that is shared across instances, within the scope of where they are declared. Thus, providing a tool for property wrapper authors to optimize their wrapper abstractions.</p> <pre><code>@propertyWrapper(shared: Storage) struct Clamped { private var value: Value struct Storage { let range: ClosedRange<Value> init(_ range: ClosedRange<Value>) { // ... } } // wrappedValue would be accessed through a subscript subscript(shared storage: Storage) -> Value { get { value } set { value = storage.range.clamping(newValue) } } } </code></pre>
ScanCode-Toolkit relies on "required phrases" to prevent false positive license detections, but 74% of its rules currently lack this protection. To solve this, I am building an AI/NLP pipeline that automatically reads raw license texts and extracts the exact phrases needed to protect these rules. I have already built a working proof-of-concept, complete with a trained model and a live web demo that the maintainers can test right now. During GSoC, I will build this into a lightweight, production ready tool. My core deliverables include a highly accurate machine learning model optimized for offline use and a native Command Line Interface (CLI). This tool will automatically inject high-confidence phrases and provide an interactive "human-in-the-loop" review mode for edge cases, ultimately delivering a measurable reduction in real world false positives.
This project aims to improve the user experience of gprMax, an open-source software for simulating electromagnetic wave propagation by modelling Ground Penetrating Radar (GPR) and electromagnetic wave propagation, through the development of an AI chatbot and assistant. The purpose of the AI chatbot is to answer and troubleshoot any questions that the gprMax users may have, while the AI assistant is capable of turning natural language into an input file in the required gprMax format. We will leverage existing pretrained LLM models and the LangChain framework, and fine-tune the LLMs using LoRA and RAG techniques, drawing on the existing gprMax documentation and years worth of discussion in the Google groups and the GitHub issue tracker for data. The ultimate goal is to deliver and deploy an AI chatbot and assistant to enhance the accessibility of gprMax, as well as save both the users and the development team valuable time by automating the troubleshooting process.
<p>This project aims at implementing new tools to support the transfer learning pipelines, ranging from object detection to embedding-based similarity matching (e.g., nearest neighbor). With these tools, users can fine-tune TensorFlow Hub (TF Hub) models with their new data directly via command lines and get their fine-tuned models saved for the ease of deployment, without writing any code. Specifically, in this proposed project, we will build a new tool for fine-tuning various object detectors provided in TF Hub, and another new tool for fine-tuning item embedding systems for structured data. Meanwhile, we will also extend the current tool for text-based nearest neighbor embedding indexing to image domains (e.g., nearest neighbor embedding indexing for images).</p>
In Bayesian statistics, one way to circumvent positing parametric assumptions is to use Dirichlet Processes (DP) as priors for probability distributions themselves. PyMC is a Python probabilistic programming library for Bayesian modelling and there are ongoing efforts to add a DP functionality to the package. DPs are daunting both theoretically and computationally as their non-trivial implementation would benefit from polished non-user facing methods in Aesara - formerly Theano - and AePPL. The former provides the computational backend to PyMC whereas AePPL converts arbitrary Aesara graphs into equivalent log-probability graphs which are convenient for automatic differentiation and hence sampling. The goals of my proposed Google Summer of Code (GSoC) project are threefold: relax the univariate assumption of components in mixture models, allow the creation of mixture graph via IfElse and Switch statements and continue ongoing efforts to add a DP functionality to PyMC.
<p>Apache Mnemonic is a non-volatile hybrid memory storage oriented library and it proposes a non-volatile/durable Java object model and durable computing service that bring several advantages to significantly improve the performance of massive real-time data processing/analytics and helps to build cache-less and SerDe-less high performance applications. Mnemonic has two memory service based on “NVML”. On december 2017 NVML has changed it’s name into PMDK and has also started growing its libraries and tools. So, we need to upgrade Mnemonic volatile and nonvolatile memory service according to PMDK libraries. PMDK is tuned and validated on both Linux and Windows, the libraries build on the DAX feature of those operating systems which allows applications to access persistent memory as memory-mapped files, as described in the SNIA NVM Programming Model.</p>
Atarashi is a modern, information-retrieval-based license scanner. Unlike traditional rule-based approaches, Atarashi leverages statistical techniques such as TF-IDF, cosine similarity, Damerau Levenshtein distance and N-gram distance to detect licenses in source code files. While the system achieves good accuracy (approximately 80%), there is still scope for improvement. This proposal aims to enhance Atarashi’s accuracy and robustness by integrating a keyword-based pre-filtering model, improving the existing classifier, and adding fallback mechanisms for ambiguous license predictions. Apart from that, from a previous initiative we have a Minerva, which is an augmented license dataset which can be used to train ML Models for licence identification. The goal is not only to refine license identification but also to integrate the work seamlessly into the current open pull request (#1634) and ensure its successful merging into the main FOSSology codebase.
<p><strong>Neural AutoRegressive Flows</strong> are one of the most recent addition to the family of autoregressive flows. By using NAFs, probability density estimation in the domain of scientific exploration can yield amazing results. For example if the background data of a particular device is modelled using NAFs, it can lead to the discovery of new phenomena with very little supervision. Main goals of the project include creating basic reusable building blocks for NAFs, <em>implementing NAFs</em> on the given High Energy Physics data obtained from various experimentation devices, <em>hyperparameter tuning</em> of the NAF model for optimum performance, providing <em>API for training plus inference</em> and finally <em>documenting</em> the API and various components of the system. Final product obtained will not only be easy to use but also easy to extend.</p>