Explore Projects

<p>This project aims to provide support for survey methods in StatsModels. In particular, survey summary statistics, adding probability weighting to existing methods such as mixed effect models and Cox models, and inference for complex surveys are primary goals. Additionally, I plan to extend GEE such that the weights are allowed to be non-constant within groups and also implement Weighted Multilevel models. I will write unit tests for all implemented approaches, testing against R and Stata.</p>

<p>I would like to work on creating dataflow operations which users can use out of the box for their data cleanup tasks. All the cleanup operations can be chained together so that users can directly plugin the data clean up operations that they would like to perform in a single dataflow, the input to that dataflow will be a dataset from some source and the output would be cleaned up data.</p> <p>I would also be creating plugin for the sklearn metrics methods for the accuracy type plugins, using which users will get many types of accuracy scorers for their models. As there are many evaluation methods of a model, using many other accuracy evaluation methods the users will be able to make an informed decision about their models.</p>

DFFML aims to democratize machine learning by providing a easy-to-use and convenient hub for the training and deployment of machine learning models. As part of this initiative, I will be implementing Automated Machine Learning (autoML) to accelerate end-user's ML workflows . To do so, I will integrate common hyperparameter tuning techniques (grid search, Bayesian optimization, evolutionary algorithms etc.) with DFFML's array of different models ranging from GBTs to NNs, and optionally ensemble learning, in order to return the best-performing model to the user. At the end of the project, users will easily be able to perform autoML with any dataset from their CLIs using DFFML, and the process will be clearly documented and tested for the benefit of both users and future contributors.

Currently, MNE-Python implements vector autoregressive modeling as a method of measuring functional connectivity across sensors used in neuroimaging. The purpose of this project will be to extend the measures of functional connectivity using a state-space modeling approach utilizing the observed variables (sensor measurements) to estimate functional connectivity across latent variables (neurological regions of interest). The output of this method will provide time-varying autoregressive (AR) coefficients, indicative of the connection strength between any two latent variables. The model’s output will be used to draw a directed graph as a visualization tool to understand information flow across latent variables. Statistical methods will be implemented in order to determine statistical differences between different network configurations.

<p>ELI5​ is a Python library for explaining and debugging machine learning (ML) models. Currently ELI5 supports scikit-learn, xgboost, and other ML libraries, taking in estimators such as linear classifiers and decision trees. However, explanations for neural networks are yet to be added. Thus, in this project I propose to add support for three popular deep learning libraries: Keras, Tensorflow, and PyTorch. I will explain the models from these libraries using an implementation of Grad-CAM​ (<a href="https://arxiv.org/abs/1610.02391" target="_blank">https://arxiv.org/abs/1610.02391</a>), a widely applicable and robust approach for highlighting what contributed to a prediction. In the case of image-based networks, a heat map will be created indicating where the network looked to make a prediction, of which there are many examples online. Less commonly done, in this project I will also explain text-based models via text highlighting.</p>

<p>The way PySAL users interact with geographic data is directly tied to manipulating unlabeled Numpy arrays. With Pandas readily available for most users, this is tedious fait accompli for users, since the core data model for PySAL's API was designed before Pandas existed. While newer projects have attempted to extend Pandas for geographic data, those packages have difficult-to-install dependencies that make them inaccessible for many end users, and make design decisions which the author thinks are somewhat suboptimal for PySAL's use case. Therefore, I propose to specify and implement a tabular spatial data model in Python leveraging Pandas dataframes directly, rather than through subclassing. This follows in the spatialite idea that spatial data is first data, then spatial. Thus, tooling must focus on enabling spatial operations directly on Pandas dataframes by exploiting column dtype information and dispatching appropriately, instead of a subclassing approach. In the end, this project would result in a simpler data model for PySAL.</p>

This project aims to build ChatPharo, an intelligent, AI-powered assistant that integrates Large Language Models (LLMs) directly into the Pharo development environment. Despite Pharo’s strengths as a live programming platform, it currently lacks a native and user-friendly interface for interacting with LLMs to assist with code explanation, generation, documentation, and debugging. To address this, the project proposes a seamless, modular, and extensible system that connects Pharo with both online (OpenAI, DeepSeek) and offline (TinyLLaMA, Mistral-7B via Ollama) models. A multi-tab chat interface will allow users to manage multiple contextual conversations, while intelligent AI commands (e.g., /explain, /generate tests, /refactor) will directly support day-to-day coding workflows. Memory optimization techniques—including smart caching, chat pruning, and lazy model loading—will ensure smooth performance even in long sessions. A key innovation includes experimenting with pseudo-parallelism in Pharo’s single-threaded architecture by spawning secondary images to offload heavy LLM processing. This enables parallel task handling without compromising the responsiveness of the main Pharo image. By midterm, users will have access to a working chat UI, multi-LLM switching, and foundational coding actions. By the final evaluation, ChatPharo will evolve into a context-aware assistant capable of inline suggestions, AI-driven debugging, and scalable architecture for future extensions. Through full documentation, tutorials, and community engagement, this project will not only enhance developer productivity but also showcase Pharo as a capable, modern AI-enhanced development platform.

<p>The project scope includes designing a web based leader board module for CMU Sphinx which would include details of top ranking people, models, training methods, data collection methodologies, proposals, benefactors, vocabulary, languages, fund raising methods, etc.</p> <p>This project would be a Python Flask based REST Web solution with inbuilt authentication which would provide users to update and view details of top performers in the above mentioned categories.</p>

<p>The proposed framework will implement a user interface capable of collecting audio signal from microphone corresponding to a prompted text and perform a comprehensive speech analysis to produce a report of mispronounced segments of speech in the collected audio. Audio collection is performed asynchronously through web workers in browsers or flash actionscript as a fallback. The collected audio is parallely decoded using multiple grammars within pocketsphinx framework and standardized acoustic scores for each decoded unit is compared across grammars to evaluate the pronunciation. These grammars form the core contribution of this project wherein they act as a panel of native and non-native transcriptionists or recognition models predicting the output as intelligible to each of them.</p>

<p>The aim of this project is to write support code for obtaining accurate transcriptions, exemplar pronunciations, and phonetic and part-of-speech labeling.</p> <p>One of the important concerns in Natural Language Processing is the availability of gold standard information such as transcripts of speech, part-of-speech tagged sentences, word pronunciation examples etc., which aid in training and evaluating the performance of speech and text processing algorithms. This requires extensive collection, processing and validation of information from a variety of sources.</p> <p>The data needs to be normalized across these sources, evaluated based on quality, quantity and other metrics. This requires a structured system which allows for the data collection and management tasks to be done effectively. This is especially important to CMU Sphinx contributors who require such data for their work, such as training and evaluating their speech and text models.</p> <p>The project involves the integration of a Flask and SQLite based leaderboard web application with a transcription record system into a broader data aggregation and processing framework for the aforementioned tasks.</p>

<p>In the past decade Deep Neural Networks (DNNs) have been be demonstrated to excel to wide range of learning tasks, including speech recognition. More recently, robust DNN-based speech recognition techniques have been developed that have been adopted by all major commercial systems including those from Google and Amazon. In such a setting, the GMM-HMM based speech recognition framework used by Sphinx is severely outdated and requires an upgrade to perform at par with the state-of-the-art systems.</p> <p>I propose a two-tiered approach for acoustic scoring that uses concepts from Convolutional Neural Networks (CNNs) for feature extraction and Gate Recurrent Units (GRU) for acoustic modelling. The first level consists of convolutional filters and pooling layers that produce convolutional features from each audio frame. The second level consists of a single GRU-based DNN that estimates the output probabilities for each HMM. The optimal state sequence can then be found using Viterbi decoding. Over the course of the project this setup will be optimised by experimenting with different structural parameters for the DNNs as well as different learning techniques.</p>

<p>This project aims to make two additions to the statsmodels package. The first is to introduce methods to handle distributed estimation of regularized and unregularized regression models. The second is to introduce several methods for statistical inference on regularized models.</p>

<p>Mapping between the natural language text of a contract and attempting to classify data types such as monetary amounts, dates and legal specific terms such as agreement parties, etc. into cicero variables from the existing model library.</p> <p>Therefore, users can export a smart legal contract by natural language contract text.</p> <h3>How to do it?</h3> <h4>A. Data Mapping using NER by RoBERTa.</h4> <p>Build a scaleable NER model by Adapter Transformers based on RoBERTa.</p> <p>The model also have Active Learning pipelines. User can define their own custom data type label then upload data and train the Adapter. By doing so, the model will recognize their new tag.</p> <h4>B. Suggest about templates by Classification Model</h4> <p>When users first upload their natural language contract, NLP model will tell them which smart legal contract template is suitable. So users can use or fine-tune the contract easily.</p> <h4>C. Identification contract variables by BERT QA Model</h4> <p>Input contract text, NLP model will suggest user which they need to put onto smart legal contract's variables.</p> <h4>D. API backend.</h4> <p>User can call the NLP model by API. Provide a Swagger UI Documents and plenty of examples.</p> <h3>For more detail, please <a href="https://github.com/accordproject/labs-cicero-classify" target="_blank">go to README in the GitHub Repo</a></h3>

<p>Most architectural and engineering projects nowadays involve a strong part of surface modeling. To do this kind of modeling objects like Bezier or NURBS curves are required. These provide a great amount of editing freedom. However, current limitations require these objects to be converted to meshes in order to perform tasks such as multiple material assignment. With this project I propose to improve curve editing by adding features that currently require the curve to be converted to mesh, or cannot be performed at all. I will also improve the documentation of the code related to curve editing (by editing documentation and by improving the comments on the code).</p>

In agent-based modeling, the goal is to understand how individual agent behavior agent behavior is influenced by social norms (rules like obligations, prohibitions, and permissions). Using platforms like CORMAS built in Pharo, modelers have to hardcode norms directly into agent behavior, making the system rigid and hard to manage. Using method proxies - a reflective programming technique in Pharo that intercepts method calls at runtime - we propose to implement norms as a separate, modular layer on top of CORMAS with different handlers for obligations, prohibitions, and permissions. This work aims to deliver 1) a CORMAS plugin (cormas-norms) with its tests and documentation, 2) a better UI to manage norms live during simulation, and 3) a real-world use case demo of the plugin.

<p>Elastic deformation is an essential component of animation pipelines. Volumetric soft body simulation can be used to capture the complex details of muscle and skin, e.g. volume preservation under collision/compression. In visual effects, such methods must also be robust to extreme deformation (no numerical blow-ups) and fast enough for interactive manipulation. Moreover, many of the underlying material models are nonlinear and computationally expensive, such as hyper-elastic models for realistic tissue or artistically-defined deformation. Toward these challenges I propose the implementation of a volumetric soft body solver in Blender, that is not only rapid and robust, but is general to hyper-elastic materials. The solver will also be capable of robustly resolving self-collisions and possibly other constraints. This would greatly improve the use of physics in Blender for animation pipelines.</p>

<p>Dipy is a python library for analysis of diffusion-weighted MRI (dMRI). Diffusion patterns can reveal microscopic details about tissue architecture and is used in clinical as well as neuroscience research. The IVIM model describes diffusion and perfusion in the signal acquired with diffusion MRI. Recently the interest has expanded and applications have emerged throughout the body including kidneys, liver, and even the heart. Many more applications are now under investigation such as imaging for cancer (prostate, liver, kidney, pancreas, etc.) and human placenta. One of its largest uses is in brain mapping and neuroscience research such as Parkinson’s disease where it is used to study aging and structural degeneration of fibre pathways in the brain. An open source implementation of this model will help researchers and students investigate such applications and will add to Dipy’s ability as a complete diffusion imaging library.</p>

<p>DFFML is a machine learning based project which provides APIs for training and testing datasets using various machine learning frameworks such as scikit-learn, tensorflow and more, making it easy to generate datasets, train and use machine learning models, and integrate machine learning into new or existing applications.</p> <p>At present, there is no way of training and testing the existing machine learning models in DFFML on image datasets, so I have selected 2 image processing python libraries OpenCV and Scikit-Image to wrap in DFFML.</p> <p>The project is divided into 2 parts:</p> <ol> <li>Wrapping the Image Processing Libraries namely OpenCV and Scikit-Image.</li> <li>High Level Operation Workflow, i.e., Custom Operations which will act as high level operations implementing a predefined flow of OpenCV and Scikit-Image functions.</li> </ol>

This project focuses on the visualization, and pre-and post-processing of opinion dynamics simulations with the Hari-Plotter code. Especially, for the application to real-world data the post- and preprocessing of data are of paramount importance. Opinion dynamics models are often simple, yet understanding patterns in potentially high-dimensional opinion spaces is a significant task. Therefore, this project aims to bridge the gap between simple toy models and complex real-world issues. Hari-Plotter is based on NetworkX to handle graphs, with several classes to handle simulations and run data. During the Google Summer of Code, I will work on handling multi-dimensional opinions, different agent types encoding various behavioral patterns and clustering opinions, and I will add functionality for averaging data over multiple simulation runs. In addition, I plan to work on adding documentation to make Hari-Plotter more user-friendly and accessible.

MNE-Python provides powerful 3D PyVista brain viewers and 2D Matplotlib sensor plots, but these tools currently operate in isolation. This project bridges that gap by extending the mne.viz.ui_events registry to allow bidirectional interactivity between source-space and sensor-space visualizations. I will implement a decoupled event system where selecting a 2D time-point updates the 3D topography, and picking a 3D vertex or ROI (Region of Interest) instantly generates the corresponding 2D evoked time-course. This solves a major UX bottleneck for researchers validating inverse solutions and interpreting neural models. Key Deliverables: - A bidirectional event-callback architecture integrated into mne.viz. - Synchronized UI updates for mne.viz.Brain and mne.Evoked.plot_* (topomaps and joint plots). - A prototype "Interactive Dipole" tool for real-time forward modeling (Xfit replacement). - Comprehensive pytest fixtures for the new event classes and a sphinx-gallery tutorial demonstrating the interactive workflow.

Transforming electrophysiology signals from sensors into their frequency representations has been shown to be a way to increase the power for application of this data; for one example of very many see Swann et al., 2015. This Fourier-transformed sensor data can be used to infer data within the brain in a very similar way as the time-series data but with an added, fifth dimension for frequency. MNE-Python has all the components necessary for time-frequency source estimate viewing and would be greatly improved by integrating them to allow for exploration of this complex data. A graphical user interface (GUI) that allows users to look at regions of high power spectral density, while letting the user dynamically adjust the range of frequencies, would be incredibly helpful for allowing this kind of data to be explored and analyzed. The ability to browse slices has already been implemented in `mne.gui.ieeg_locate` and could be abstracted to be used in this case. Time-frequency plotting of spectrograms, with appropriate units and baseline correction already exists in MNE-Python as well. Putting these elements together into a GUI would be a great improvement for the MNE-Python tool and community.

<p>Currently <strong>pgmpy</strong> provides two sampling classes, A class of algorithms namely Forward sampling, Rejection Sampling and Likelihood weighted sampling which are specific to Bayesian Model( BayesianModel in pgmpy) and Gibbs Sampling a Markov Chain Monte Carlo Algorithm that generates samples from both Bayesian Network and Markov models. Since pgmpy is looking to support continuous random variables in coming time, we need classes of inference algorithms (sampling too) that can specifically work with these continuous random variables. This proposal deals with adding two more sampling algorithms in pgmpy, namely:</p> <ul> <li><strong>Hamiltonian/Hybrid Monte Carlo (HMC)</strong>: A Markov Chain Monte Carlo algorithm that adopts physical system dynamics rather than a probability distribution to propose future states in the Markov chain.</li> <li><strong>No U Turn Sampler (NUTS)</strong>: An extension of Hamiltonian Monte Carlo that does not require the number of steps L (a parameter that is crucial for good performance in case of HMC).</li> </ul>

To build a DJ application with two decks in Pharo, I will make use of three libraries: Phausto for audio file playback (at variable speed), equalization, and filtering. Spec for building a browser to visualize audio files in a folder together with their metadata, and to organize these files into playlists. Bloc for the visualization of audio waveforms and the rendering of UI widgets to control playback speed, pitch, levels, the equalizer, and filter cutoff and resonance. As a first step, Phausto will be used to load and simultaneously play two audio files, mixing them via the Phausto API. Once the audio engine is in place, a Spec application will be built to browse audio files and load them into either of the two decks. With the file browser complete, a conventional user interface will be added to control the mix between the two tracks and their equalization. The Phausto standard library will then be used to add a delay effect per deck. With the DJ application running, final polish can be applied in the form of BPM detection and track quantization, which requires implementing peak detection for audio files within Pharo. Deliverables: The Spec application with audio file browsing and audio playback will be delivered before the first evaluation. The audio visualization and UI built with Bloc will be delivered halfway between the first and final evaluations. Extra features (BPM detection and quantization) will be added and tested before the final evaluation.