Explore Projects

This project aims to significantly enhance Mesa-Frames by building efficient and scalable data collection tools for large-scale agent-based simulations. The core focus is on developing a flexible framework that allows researchers to analyze simulation data with greater precision and lower overhead. One of the main components, the Stats Collection, will enable users to define exactly which statistics they want to track—such as mean, max, or count—helping reduce both memory usage and computation time. Another key feature is the Event Driven Collection, which records data only when predefined conditions are met. This ensures researchers capture meaningful insights while avoiding unnecessary logging and storage. To support even larger simulations, the project will also include integration with external storage systems like PostgreSQL and S3, allowing collected data to be streamed directly to disk or the cloud. In addition to these core improvements, the project will focus on performance optimization through lazy execution with Polars and better vectorization. Clear, modular documentation and a focus on usability will ensure Mesa-Frames remains an accessible and powerful tool.

Mesa offers a suite of multi-agent environments tailored to diverse objectives. These environments find natural applications in Multi-Agent Reinforcement Learning (MARL). The project aims to empower users further by leveraging Mesa for MARL endeavours. A possible extension is desired which could facilitate seamless integration with popular Python reinforcement learning libraries such as Stable Baselines and RLlib.

As ABMs are simulations and often have phase transitions (periods of rapid change to new stable states), being able to go back in time and replay key results would be a great addition to Mesa. Critically, no computation would be needed as the results are stored. This project aims to implement caching for Mesa, with options for users to select the critical results, so that the critical parts of the simulation can be replayed. For long simulations, they can access interesting points in the simulation quickly and easily.

The goal of this project is to optimize the FLINT contributor's workflow by automating the complex difficult steps , providing user-friendly feedback that is easy to share.I plan to standardise the workflow by setting some automated procedures that utilize some standard MLOps techniques (e.g. DVC).

This project aims to build a faster, lightweight, user-friendly, mobile-friendly, and cross-browser-compatible UI library as a stand-alone dependency for moja global projects. Moja global User-interface (UI) library project aims to bring forward an intuitive, consistent, and easy-to-use interface that can help the developers within the User-Interface working group and users to quickly accomplish their tasks. The UI library aims to considerably improve the design & development workflow and meet the acceptable web accessibility requirements for potential users. A UI library helps to mitigate popular UI-development issues like inconsistent user experience, performance issues, accessibility requirements, and more. To achieve this, a centralized collection of components would be developed which will encompass the color branding of moja global, typography, spacing, buttons, modals, and forms. The big picture is that the UI library will allow all the upcoming moja global projects, with a client-facing focus, to have a consistent design language that will help us to quickly and easily build their user interface.

Publish fine-tuned MobileViT in TensorFlow Hub TensorFlow Hub is the main TensorFlow model repository with thousands of pre-trained models with documentation, sample code and readily available to use or fine-tune. The idea behind the project is to develop new State-of-the-Art models like MobileViT and publish the pre-trained models on TensorFlow Hub using the ImageNet1k dataset. MobileViT is light-weight and general-purpose vision transformer for mobile devices. MobileViT presents a different perspective for the global processing of information with transformers, i.e., transformers as convolutions. Our results show that MobileViT significantly outperforms CNN- and ViT-based networks across different tasks and datasets. On the ImageNet-1k dataset, MobileViT achieves top-1 accuracy of 78.4% with about 6 million parameters, which is 3.2% and 6.2% more accurate than MobileNetv3 (CNN-based) and DeIT (ViT-based) for a similar number of parameters. On the MS-COCO object detection task, MobileViT is 5.7% more accurate than MobileNetv3 for a similar number of parameters.

My project is a development of 3 computer vision models: 1. A model for workflow classification; 2. A model for medical equipment detection; 3. A model for scene segmentation (segmenting organs and surgical Tools); NB: I have found out that there is a lack of VR medical imaginary datasets. That is why I suggest to use domain adaptation technique for developing ML models. In other words, I suggest to compare two approach: transforming ordinary images to make them similar to VR images, transforming VR images to adopt them to the original data domain.

In GSoC 2025, I prepared a novel physics-informed Swin Transformer to classify between lens models in simulated gravitational lensing images. The model overtook its contemporaries in sparse datasets. However, the core challenge often is that models trained on perfect simulation data often fail when applied to noisy, complex, and unlabeled observational data, like the Legacy Survey of Space and Time corpus. Thus, I am proposing an application to that model, through a real-life LSST dataset. The strategy that will be used will be a hybrid of “conventional” substructure classification using vision transformers, and a masked autoencoder, as the dataset itself is largely unlabeled. This will allow us to apply the novel architecture defined in the last project into real-world unlabeled datasets like the LSST corpus. The deliverable will be a foundational model with its core being the HEAL-Swin model, having both the strengths of a a physics-informed relativistic inference, and applicability to real-world unlabeled datasets. The aim will be to deliver publishable results by the end of the program.

This project focuses on adding full 3MF export and import support to InVesalius, enabling efficient handling of multi-surface medical models with accurate color (RGBA) representation and proper coordinate transformations. At present, InVesalius does not support modern 3D formats like 3MF, which offer better structure, metadata handling, and color capabilities compared to formats such as STL and OBJ. Through this project, I aim to integrate 3MF support into the existing export and import pipeline using lib3mf, ensuring compatibility with tools like OrcaSlicer while preserving geometry, scaling, and visual attributes. The implementation will handle coordinate transformations (LPS ↔ Z-up), support multi-surface models, include alpha channel handling, and integrate cleanly with the current architecture. This work builds on my prior contributions to InVesalius, where I have worked on surface pipelines, coordinate transformations, and UI workflows, allowing me to integrate this feature effectively within the existing system. The goal is to improve interoperability with modern 3D tools, ensure accurate representation of models, and provide a solid foundation for future extensions in 3D medical imaging workflows.

This proposal aims at developing AI-generated choreography for duets. It outlines a novel approach leveraging both Transformer and Graph Neural Networks to analyze and generate dance movements, focusing on creating authentic dance material that reflect the intricate dynamics between dancers. The project will use Transformer and/or GNN to process motion capture data. The comprehensive solution plan involves: · Understanding key relationships between parts of the body of each dancer through a dedicated Spatial/Temporal Joint Relation Module · Utilizing self-attention and cross-attention mechanisms to learn spatial-temporal relationships of dancers' movements · Enhancing the overall coherence of the generated choreography by implementing regression loss and style similarity loss · Create a dataset of point-cloud data corresponding to extracted motion capture poses from videos of dance duets · Utilize Transformer and/or GNN models with encoder-decoder structure and various attention mechanisms to generate the movements of Dancer #2 conditioned on the inputs of Dancer #1 and/or invent new, physically-plausible duet phrases · Implement a Spatial/Temporal Joint Relation Module to learn key relationships between parts of the body of each dancer that are integral to the dynamics of the duet · If time permits: incorporate Music Embedding Module to enrich the model with multi-modal inputs · Collaborate with the original dancers to use the model outputs to inspire new performance material

Accurate gene annotation in eukaryotes solely based on genomic data has been a significant obstacle in biology since the introduction of next-generation sequencing technologies and thus the rapid increase of available data. Traditional methods either rely on homology searches to map the open reading frames to previously identified protein-coding genes or utilize additional experimental data, e.g., transcriptomics data. The first approach produces potentially inaccurate results if the genome of interest is not at least somewhat related to an already annotated genome. The second approach is hindered because gathering transcriptomic data is labor-intensive and expensive. For that reason, there is a high demand for models that predict the location of protein-coding genes solely from inherent features of the DNA sequence of the gene. Although theoretically possible, methods that use, for instance, Hidden Markov models to detect protein-coding genes based on known gene features are often inaccurate. In this project, we will train a Deep Learning Transformer model to extract features of protein-coding genes to gain deeper insight into their exact properties that lead to translation. The whole workflow will include first training a Conditional Random Field model to recognize candidate gene regions and then using these as input for a more fine-grained Transformer - Convolutional Neural Network hybrid model. The final pipeline will be tested against a benchmark of gold standard annotations as well as various test sets to evaluate the influence of different parameters like genome sequence quality, protein length or gene structure complexity.

The torchvision package in the mlverse ecosystem provides essential tools for computer vision workflows in R, including datasets, transformations, and model utilities. While the ecosystem has matured significantly, a key limitation remains in the performance of certain core operations, particularly in detection pipelines. Functions such as Non-Maximum Suppression, bounding box transformations, and proposal filtering are currently implemented at the R level, which introduces inefficiencies when working with larger inputs or real-time use cases. This project aims to address that gap by moving performance-critical components into a native C++ backend using torchvisionlib. The approach will focus on identifying the parts of the pipeline that are actual bottlenecks, implementing them in C++ with careful attention to memory usage and batching, and integrating them back into torchvision without changing the user-facing API. Rather than rewriting everything, the work will prioritise targeted optimisations that have a measurable impact on end-to-end performance. In addition to backend improvements, the project will include the implementation of at least one modern computer vision model family, such as YOLO, using the optimised components. If time permits, support for an additional model (e.g., RT-DETR) will also be explored. These models will be integrated into the existing ecosystem with consistent APIs and working examples. The main deliverables of this project are: a set of optimised C++ implementations for core detection operations, their seamless integration into torchvision, at least one fully functional model pipeline built on top of this backend, and supporting documentation and benchmarks demonstrating performance improvements. The goal is to make torchvision not just feature-complete, but efficient enough for practical, large-scale use.

I propose to contribute to the OpenClimateFix organization by implementing the GenCast model (https://arxiv.org/abs/2312.15796), with a focus on advancing weather forecasting accuracy. By integrating state-of-the-art techniques such as diffusion models, sparse transformers, and graph neural networks, this project aims to address the challenges of improving weather prediction and contributing to climate research. The key deliverables of this project include: - Dataset Integration: incorporating the ERA5 dataset from WeatherBench2 to make comparison between different models easier. - Graph Mesh and Noise Generation: constructing the graph mesh and generating isotropic noise. - Model Architecture Enhancement: modifying the GraphWeather architecture, particularly focusing on the integration of sparse transformers in the Processor. - Conditional Layer-Norm Implementation: replacing standard layer normalization with conditional layer normalization. - Loss Function Adaptation: adjusting the loss function to incorporate noise level-dependent weighting. - Sampling: integrating the second-order DPMSolver++2S sampler for efficient score-based forecasts generation. - Testing and Documentation: adding tests during every phase of the project.

<p>Data-driven background estimation is crucial for many scientific searches, including searches for new phenomena in experimental datasets. Neural autoregressive flows (NAF) is a deep generative model that can be used for general transformations and is therefore attractive for this application. The MLBENDER project focuses on studying how to develop such transformations that can be learned and applied to a region of interest. In this project the main aim is implementing a Neural Autoregressive Flow (NAF) model to estimate the background distribution and apply it to a representative physics analysis searching for a resonance excess over a smooth background.</p>

<p>The goal of this project is to develop a model that can visualize biological networks as logic gate circuits, and vice-versa. In recent years, discrete dynamic modeling has emerged as a powerful method of modeling and understanding biological systems. One type, boolean network models (which was initially used as a prototypical model of genetic regulatory networks), has been shown to be a tractable approach for modeling large-systems. A boolean network consists of a set of nodes whose state is binary and is determined only by other nodes in the network through Boolean functions. The simplicity and self-containment of these graphs makes it extremely fast to simulate these systems and is also a very realistic interpretation of how biological networks actually work. Unfortunately, no libraries yet exist for converting a visualization of a biological network into a logic gate visualization and vice-versa.</p> <p>Thus, the aim of this GSoC project is to develop a web-based tool that can visualize biological networks, and also transform these networks into a logic gate-based system.</p>

<p>DeepChem has been great library for the application of deep learning for drug/chemical discovery. As of now there is no imaging models implemented in DeepChem that would facilitate the use of medical data (like images of brain scans, or UltraSound images etc). This project proposes to build an API for data augmentation for imaging which in recent times has shown to make models invariant slight data transformations like rotation, translation, noise etc. and as well as to build 2 models : one being the U-Net for bio-medical image segmentation and the other being the ResNet-50 model (trained on the imagenet), that would allow users/researchers to use these pre-trained models and extend it to implement their own networks with much ease.</p>

The program aims to leverage agents built using publicly available large language models (LLMs) and personalize them with Retrieval-Augmented Generation (RAG). This approach creates various levels of agentic roles to transform mainframe workloads and modernize them at every phase, from discovery and design to transformation, testing, and deployment. Problems to be addressed in the project: - Unavailability of application documentation. - Lack of comprehensive view of all tasks to be included in the transformation plan. - Multiple iterations of Code conversions. - Technical and resources dependencies. - Unavailability of Application support for the transformation. Deliverables: - Working Agents which will build a program plan at each of the modernization phase for all workstreams. - Agents performing Code transformation with minimal effort from migration team. - Agents performing application analysis with less dependency from application teams.

The "Superresolution for Strong Gravitational Lensing" project aims to enhance the resolution of astronomical images through advanced deep learning techniques. Gravitational lensing magnifies distant galaxies, offering insights into the universe's dark matter structure. However, the limited resolution of these images constrains our understanding. This project will develop superresolution models based on conditional diffusion models to transform low-resolution gravitational lensing images into high-resolution counterparts. These models will be tailored for astronomical data's unique challenges and integrated into the DeepLense framework for broader scientific use. Deliverables include a prototype superresolution model, integration tools and documentation for DeepLense, and comprehensive model development documentation. This effort will improve gravitational lensing image quality, aiding in the study of cosmic structures and dark matter.

In high-energy physics (HEP), automating the symbolic computation of squared amplitudes is essential for predicting cross-sections and validating theoretical models against experimental data. Current transformer-based approaches excel in sequence modeling but falter in capturing the relational structures and physical consistency required for HEP expressions, such as those involving gamma matrices and Mandelstam variables. To tackle this, I propose the Physics-Informed Performer (PIP), a novel framework that integrates the Performer transformer’s efficient kernel-based attention with a physics-informed tokenization scheme and a SymPy-based validation layer. This approach leverages Performer’s ability to handle long sequences, introduces tokens like <SQUARE> and <S> to embed HEP-specific operations and kinematics, and ensures symbolic accuracy through empirical post-processing. By focusing on data representation, PIP offers a computationally efficient alternative to traditional symbolic methods, tailored to the Task 1.2 dataset of amplitude-to-squared-amplitude mappings.

This project addresses the challenge of integrating computationally intensive Machine Learning inference into responsive Scala applications without compromising service latency. Currently, ML inference within Cats Effect applications can monopolize CPU resources, starving concurrent operations. The proposed solution develops a compiler that transforms standard ONNX machine learning models into optimized Cats Effect IO operations using Scala Native for high-performance numerical computation via mlpack C++ bindings. The project will deliver: (1) C bindings for mlpack with Scala Native integration, (2) a custom Intermediate Representation for ML operations, (3) an ONNX parser, (4) a core inference library that transforms ML models into Cats Effect IO actions with strategic yielding, (5) a functional MNIST digit recognition web service demo with benchmarking capabilities, and (6) comprehensive documentation and tests. This solution will enable seamless, efficient, and purely functional ML integration within the Typelevel ecosystem without degrading service responsiveness

The ATLAS Tile Calorimeter (TileCal) currently reconstructs energy using a linear signal-processing method. For the Phase-II HL-LHC upgrade, readout will move to a 9-sample sliding window, where this approach becomes less reliable because many proton-proton collisions overlap in time. This project develops compact neural networks (MLPs and 1D-CNNs) that improve energy reconstruction while remaining small enough for real-time FPGA deployment. The approach combines a floating-point teacher CNN trained offline, knowledge distillation into a compact student model, quantization-aware training with 1-8 bit mixed precision, and magnitude pruning, targeting a strict 225 ns algorithmic latency within the 1.7 μs TilePPr budget. Deliverables include reproduced linear baselines on Phase-II simulation data, a trained teacher CNN with benchmarked energy resolution, a distilled and quantized student model, a prototype offline reconstruction integration, and a full FPGA synthesis report with latency and resource usage on a Xilinx Ultrascale+ target device.

This project will enhance FOSSology's license intelligence stack by revamping both Nirjas and Atarashi into a faster, ML-driven, and language-aware pipeline. Nirjas will move from regex-oriented parsing to a Tree-sitter architecture with language packs across 248 languages, improving comment extraction fidelity in real repositories. Atarashi will be upgraded from broad full-file matching toward retrieval-plus-classification that emphasizes high-signal fragments, because keywords matter for license intent (for example, phrases like permission grant, redistribution, warranty disclaimer, copyleft, and patent terms are often decisive). To support both components, I have already upgraded the Minerva dataset pipeline to generate train/validation/test datasets for Nirjas and Atarashi using merged inputs from ScanCode LicenseDB, FOSSology licenseRef.json endpoint, and The Stack Smol comment corpus, followed by augmentation, LLM-generated hard-negative synthesis, and near-dedup. A sample dataset for benchmarking and integration testing is ready, while the full-scale dataset build is in progress. For modeling, Nirjas will use a model2vec classifier distilled from strong sentence-transformer teachers selected using MTEB/RTEB task performance together with our internal evaluations; the same model2vec distillation path is also planned for Atarashi, with a fine-tuned quantized embedding baseline retained as a fallback depending on speed/accuracy trade-offs. The expected outcome is a production-ready path to more accurate, scalable, and maintainable license detection across diverse codebases.

Geant4 is a highly accurate and detailed simulation toolkit used for simulating the passage of particles through matter. Due to its strict precision requirements, simulation is slow and has been proven to be a bottleneck for physics analysis. To overcome this bottleneck, popular machine learning techniques like generative modelling have been employed as a fast simulation alternative. This project focused on the Optimization and Inference components of the project. Very limited work is done on the model in terms of Memory Footprint Reduction. This project was aimed at implementing different post training Memory Footprint Reduction techniques by creating a pipeline using KubeFlow, consolidating all the insights from different experiments and use them to extend KubeFlow pipeline built for carrying out the experiments into a more generalized form which can be used on different ML models at CERN to output the best optimized model. I integrated different Onnx Runtime Execution Providers - MLAS (default CPU), CUDA, TensorRT, oneDNN and OpenVINO into the Geant4 Par04 inference module. I have implemented an end-to-end KubeFlow pipeline which contains CPU and CUDA optimization modules which leverage different quantization and graph optimization workflows and reduce memory footprint of models while maintaining similar performance. Optimization modules for oneDNN and TensorRT are currently in-development at the time when GSoC is ending.