Explore Projects

The Machine Learning Scorecard Phase - 5 is a project listed as a part of the GSoC’22 program. The project involves making significant improvements to the existing loan and credit scoring system by exploiting the use of machine learning, pipelines, and modern statistical models. It also aims to bring in the feature of credit bureau integration which can significantly help the accuracy and minimize losses to financial institutions while doling out credit to customers. With the completion of this project, the credit scoring system in the fineract suite will be transformed into a very strong part of the loan application module.

In Google Summer of Code 2024, I focused on optimizing key algorithms in the mlpack library, targeting both performance and memory usage across neural networks, clustering methods (KMeans, DBScan), and SVD. Highlights of the project include implementing FFT for faster convolution operations in neural networks, applying post-training quantization to reduce model sizes, and parallelizing KMeans and DBScan with OpenMP for improved execution speed. Additionally, I optimized memory usage and reduced binary sizes by refining the code structure. These enhancements have made mlpack faster and more efficient, contributing to its effectiveness for a wide range of machine learning applications.

<p>My proposal involves developing an ideal model for sentiment analysis for OpenNLP by starting with a simple model and advancing to the desired advanced model applicable in many situations. I will be using Twitter data for my tests, and my final outcome will include the final developed advanced model specified in the article and documentation describing how it was created, how it works, how efficient and accurate it is, as well as tutorial on how to learn it. I have already started working on some parts of the project.</p>

<p>Bayesian Additive Regression Trees (BART) is a Bayesian nonparametric approach to estimating functions using regression trees. A BART model consist on a sum of regression trees with (homoskedastic) normal additive noise. Regression trees are defined by recursively partitioning the input space, and defining a local model in each resulting region of input space in order to approximate some unknown function. BARTs are useful and flexible model to capture interactions and non-linearities and have been proved useful tools for variable selection.</p> <p>Bayesian Additive Regression Trees will allow PyMC3 users to perform regressions with a “canned” non-parametric model. By simple calling a method, users will obtain the mean regressor plus the uncertainty estimation in a fully Bayesian way. This can be used later to predict on hold-out data. Furthermore, the implemented BART model will allow experience users to specify their own priors for the specific problem they are tackling, improving performance substantially.</p>

<p>In the last decade simulation models have become more prevalent in a variety of application areas. Each of these models define a set of input parameters which determine the features of the output. With enough of these outputs we can begin to perform what is known as reconstructing the black box in the field of parameter space analysis (Konyha et al.). One method to do this is through trial and error, where a researcher sets the input parameters to specific values based upon their experience, and collects the outputs as sample points to investigate later on (Sedlmair et al.). This is a computationally expensive method for exploring parameter space which limits the effectiveness of the program overall.</p> <p>This project represents an effort to update the Parameter Space Exploration (PSE) tool so that it can be used effectively for exploring the parameter space associated with each included large-scale brain model. This achieved would effectively reconstruct the black box of the model which is vital for our understanding of its validity, and supports ongoing work to better understand the original basis for our model, the human brain.</p>

SAIL-RISC-V golden model is currently being actively developed and focuses on the ISS export feature whereby a reference interpreter/emulator is automatically generated from the specification. Prior work on RGEN has shown that export the LLVM (as well as QEMU and PyTorch) is possible, to support automatic flow of custom instruction extensions (modules) to compiler intrinsics, emulators and AI frameworks. In this project, the export feature shall be extended to cover CGEN, the compiler ISA definition/extension framework for GCC. Building on initial work done at Embecosm earlier, we will upgrade the CGEN framework, build transformations to automatically generate CGEN descriptions given a SAIL specification using an example of tensor/matrix multiply extension.

Training intricate machine learning models to achieve the desired level of proficiency can be time-consuming and require high computational resources. These challenges restrict the utilization of highly complex models in mlpack. To tackle this issue, one solution is to utilize pre-trained models available in the ONNX format. ONNX offers a standardized method for developers to define and exchange deep learning models across various frameworks like TensorFlow, PyTorch, and more.

Red Hen Lab focuses on various aspects of linguistics and cognitive science, including Construction Grammar and FrameNet. With a voluminous collection of research publications, Red Hen Lab receives numerous inquiries from interested individuals seeking guidance. However, due to resource constraints, they are unable to respond comprehensively to all queries. To address this issue, the proposal aims to develop two open-source AI chatbots capable of engaging with users focusing on Red Hen Lab itself as well as on Construction Grammar and FrameNet. This document proposes the utilization of a transformer-based generative model augmented with Retrieval-Augmented Generation and meticulously fine-tuned on domain-specific data sourced from Red Hen Lab and FrameNet/Construction Grammar.

ArviZ's purpose is to provide backend-agnostic tools for exploratory analysis of Bayesian models. To do so, we transform data from probabilistic programming frameworks to InferenceData, which we can then use in our plotting. Besides removing the need to explicitly import ArviZ with PyCall.jl, ArviZ.jl extends ArviZ with functionality for converting Julia types into ArviZ's InferenceData format. It also allows smoother usage with PyPlot.jl and provides functions that can be overloaded by other packages to enable their types to be used with ArviZ. Gen.jl is a general-purpose probabilistic programming system with programmable inference, embedded in Julia. Arviz currently isn’t compatible with Gen.jl and our goal is to implement a converter similar to the ‘from_xyz’ methods that are currently available in Arviz,jl.

In today's data-driven world, leveraging legacy data can be a game-changer for businesses. By integrating AI with Retrieval-Augmented Generation (RAG), organizations can unlock the hidden potential of their historical data. RAG combines the power of large language models with a retrieval mechanism that fetches relevant information from vast datasets. This approach not only enhances the accuracy and relevance of AI-generated responses but also ensures that valuable insights from legacy data are utilized effectively. Whether it's improving customer service, optimizing operations, or driving innovation, RAG empowers businesses to transform their legacy data into a rich source of actionable intelligence, paving the way for a more informed and strategic future. Deliverables: Detailed research and conceptual framework document (whitepaper) to implement AI that uses mainframe data.

This project addresses the inverse problem of estimating physiological parameters that characterize the permeability of the blood-brain barrier (BBB) from MRI data. These parameters cannot be measured directly and must be inferred from signal progression over time in techniques such as Dynamic Contrast-Enhanced MRI (DCE-MRI) and Arterial Spin Labeling (ASL). The estimation is complicated by measurement noise, model ambiguity, and the sensitivity of parameter estimates to small variations in the data. To improve robustness and accuracy, this project will implement a Bayesian framework that incorporates prior physiological knowledge and observed MRI signals to estimate the posterior distribution of the parameters. For the pharmacokinetic modeling the Tofts model for DCE-MRI and perfusion modeling for ASL will be used. Markov Chain (MCMC) sampling and gradient-based optimization methods may improve the robustness of the reconstruction. Deliverables include: --- An extensible implementation of the BBB modeling pipeline using OSIPI tools --- Support for both DCE-MRI and ASL data --- Parameter estimation via Bayesian inference and optimization --- Comprehensive documentation and integration into the OSIPI project.

<p>The GoPro Cineform codec is a cross-platform intermediate codec designed for editing high-resolution footage. It was original designed for compressed Digital Intermediate workflows for film or television applications using HD or higher resolution media. FFmpeg reverse engineered Cineform decoder currently supports progressive video files with image formats YUV 4:2:2 10-bit, RGB and RGBA 4:4:4 12-bit. This project aims to support more features present in cineform codec, namely Inverse Temporal Transform for p-frames, deinterlacing support, assembly instructions integration into the ffmpeg version, and making it faster by using threading models present in FFmpeg. The result of this project will be that Cineform decoder can be integrated with many tools that use FFmpeg libav.</p>

The radiology Artifact Detection aims to create a system that identifies image artifacts on user selected regions of interest. till now the radiologists are manually finding the artifacts . in this project i aim to create a system that helps the radiologists in this process . The scanning reports will contain 100's of scripts and the traditional segmentation models works on pre trained data and breakdowns when it encounters a rare or a completely new data. The solution which i have planned is using an MedSAM model as it classify the scanning reports based on the visual signatures of the user's selection. A 2.5D volumetric tracking in this once an artifact is identified on Slice N, the algorithm prioritizes searching the exact spatial coordinates on Slices N-1 and N+1 which reduces the requirement of computation power. planning of using the human in the loop method in which out of 10 if the 8 predictions are correct and 2 are wrong if the user deletes 2 wrong predictions then the model learns that these two are wrong. i planed of using DICOM for datamanagement. The structure: Frontend -> model -> datainput -> MedSAM model -> loss calculation and scoring -> datamanagement.

Cantor is a core tool in the KDE ecosystem that focuses on scientific computing and interactive analysis. It supports mathematical modeling and data visualization through engines such as Python, Julia, and Maxima. In order to improve its text editing experience and maintainability, this project plans to deeply integrate the KTextEditor framework to transform Cantor into a more "developer-friendly" scientific research platform. By using QGraphicsProxyWidget to seamlessly replace existing custom cells, professional functions such as Vi editing mode, intelligent code completion, dynamic syntax highlighting (supporting user-defined symbol injection) are implemented, while simplifying the code structure and reducing redundant component code. This improvement not only aligns Cantor's editing experience with KDE professional tools such as Kate and KDevelop, but also lays the foundation for future integrated debuggers, version control and other advanced functions by sharing the KTextEditor plug-in ecosystem.

This project transforms API Dash into a next-generation, AI-native API client. I am merging three core objectives: 1. Agentic API Testing, 2. MCP Support 3. Generative UI Visualizations into a unified Dual-Transport Model Context Protocol (MCP) Architecture: 1. Agentic Sandbox & Generative UI (GenUI): Eliminating brittle WebViews, the internal UI will utilize Google's genui SDK and the Open Responses specification. This powers two features: an autonomous, self-healing Agentic Testing Dashboard, and a "Generative UI Previewer" that visualizes rich AI responses for end-users building Flutter/Web apps. 2. Headless MCP Engine: This exposes API Dash as a standard stdio MCP Server for external Agent interfaces (like VS Code, Cursor, or Claude Desktop).

The project, "Data Processing Pipeline for the LSST," is designed to bridge the gap between the LSST data ecosystem and DeepLense's deep learning workflows. The proposal outlines the development of a modular, efficient pipeline that automatically retrieves relevant LSST images and catalogues via the official LSST Butler API, applies configurable preprocessing (such as cutout extraction, normalization, resizing, and channel management), and finally delivers analysis-ready data into DeepLense ML models for tasks like lens detection, classification, and super-resolution. Key deliverables include: 1. A robust LSST Data Access Module that reliably fetches data. 2. A flexible preprocessing module for transforming raw images into ML-compatible formats. 3. Seamless integration with DeepLense’s deep learning workflows, enabling end-to-end processing. 4. Comprehensive documentation and reusable, well-structured code.

Machine learning techniques are regarded to have the potential to help researchers better comprehend dark matter. Convolutional Neural Networks (CNNs) have previously been used to successfully distinguish between distinct types of substructures such as no substructure, vortex substructure, and spherical substructure. In this task, we'll classify images based on whether or not they have a strong lensing effect. Strong Gravitational Lensing allows us to analyse the substructure of dark matter, allowing us to gain a deeper understanding of its underlying nature. The observed strong lens images may be utilized in the Deeplense pipeline for dark matter substructure classification, anomaly detection, and interpretation. Images may be processed using a variety of algorithms. We'll try out several types of models, fine-tune them to obtain good outcomes, and compare them. Because a single model may not be sufficient, we can combine the findings of several models using ensembling to obtain a more robust and better outcome. We'll document and record every outcome as we experiment with our models. We will also provide starter notebooks describing how to train and test the models so that it will be easy for anyone to understand our work.

<p>MediaPipe fills the gap of handling the implementation of models for different systems and allows the developers to focus more on experimentation than on system architectures. WebAssembly has become prominent in recent years as it can port C++ applications onto web browsers, making code more portable and utilizing the efficiencies of C++. In my GSoC internship, I will combine MediaPipe and WebAssembly by building an open-source realtime web video effects app. I will create a complete workflow of wrapping models in MediaPipe to be used in a webapp for running different ML-based video effects live in the browser using WebAssembly. The workflow will include wrapping the model binaries in calculators, create calculator graphs to use the model binaries, set up Bazel for WebAssembly compilation, write Emscripten bindings to access the models, and create a Typescript webapp to demonstrate how to access the WebAssembly binaries in realtime. I have extensive experience in C++ development, along with web development using Javascript. I have also been using MediaPipe for almost a year and am accustomed to the codebase. Hence, I am confident that I will be successful in my internship.</p>

This project aims to transform how analysts interact with threat intelligence data in IntelOwl by introducing a conversational AI interface. Currently, analysts must manually navigate dashboards, query APIs, and correlate results across different jobs and investigations. To solve this, I will integrate a privacy-preserving, self-hosted LLM chatbot using Ollama and a LangChain ReAct agent. The agent will be equipped with custom tools to query IntelOwl's internal REST API using natural language. Key deliverables include: 1) A new Django chatbot app handling WebSocket connections via Django Channels for real-time token streaming. 2) Ollama Docker integration with CPU/GPU auto-detection. 3) A native React chat component embedded in the existing frontend. 4) Custom LLM tools for triggering analyses and searching data models. 5) Comprehensive test coverage and detailed documentation for users and future maintainers.

<p>The package <code>mlr</code> is a comprehensive machine learning toolkit for R, providing a standardized interface to over sixty machine learning R packages, in combination with a wide range of features related to visualization, data manipulation, model evaluation and selection, and parameter tuning. Even though mlr offers the possibility of performing automatic data preprocessing when applying a machine learning algorithm, the current implementation is relatively limited in scope and functionality. This project seeks to extend mlr's capability in this regard, by developing a supplementary package <code>mlrCPO</code> with an API that gives more flexibility to the user, and by providing access to a wider range of preprocessing methods. The project introduces a first-class <code>CPO</code> ("Composable Preprocessing Operator") object that represents a particular data transformation procedure, and which can be organized in pipelines using a composition operator <code>%&gt;&gt;%</code>. Many <code>CPO</code> classes implementing the most popular and widely used preprocessing methods are implemented.</p>

The LHC at CERN generates 30 petabytes of physics data every year. Current compression tools like LZMA treat this data like any other file and miss the statistical patterns unique to particle physics. The BOA Constrictor solves this by training a neural network to predict the next byte more accurately inside an entropy coder. The original BOA paper proved the concept works (4.03x compression on CMS data vs LZMA's 3.22x), but two gaps remain: Mamba doesn't install on Python 3.12, and nobody has compared alternative architectures under fair conditions. This project will build a benchmark suite comparing GRU, LSTM, MinGRU, Transformer, and Mamba under parameter-matched conditions. It will then optimize the best backbone using INT8 quantisation, pruning, and distillation, producing Pareto frontier plots showing compression ratio vs model size. Finally, it will explore physics-informed priors as an auxiliary loss to improve predictions. All code and results will be contributed back to the upstream BOA repository.

The M-Competitions are a series of open competitions to evaluate forecasting methods for time-series data. The most recent competition in this series, M5, had their top submissions as Deep Learning models. Looking at these advances, the implementation of sktime’s deep learning module leaves a lot to be desired. As part of this project, the following will be the critical additions to the deep learning module: Add simple Deep Learning models for Forecasting like LSTMs, GRUs, MLP, etc. Include versatile and best performing models used in M5 for Forecasting (like N-Beats, use of LightGBM) Research existing best performing forecasting and classification models and implement them if required (e.g., methods based on Transfer Learning) Add soft functionality for ease of use (ensemble methods, send dataloader as input etc.) Provide improved documentation and tutorial notebooks for ease of understanding of users. After this project, users will be able to use deep learning models for forecasting and classification problems. They will be able to conveniently integrate sktime with their code, with an improved interface, varied inputs, better documentation, and tutorials for usage.

<p>The increasing number of sensors orbiting the earth is systematically producing larger volumes of data, with better spatiotemporal resolutions. To deal with that, better accurate machine learning approaches, such as Deep Learning (DL), are needed to transform raw data into applicable Information. Several DL architectures (e.g. CNN, semantic segmentation) rely only at spatial dimension to perform, for example, land-cover/land-use (LCLU) maps, disregarding the temporal dependencies between pixels observations over the time. Also, high-res remote sensing data (e.g. Planet, Sentinel) may provide more consistent time-series, that can be use in the identification of important LCLU classes, like crop, pastureland and grasslands.</p> <p>This potential can be explored using Recurrent Neural Networks (RNN), a specific family of DL approaches which can take into account time dimension. A promising project idea would be implement a RNN approach (e.g. LSTM) to classify a Sentinel time-series that is able to produce a map for one or more LCLU classes and ship the final version of the model into the Pycoal library as part of the COAL project.</p>