
NumFOCUS promotes open source scientific software.
Showing 5 of 200 projects. Click any project card for scope, mentors, and proposal studio.
Mentors: Student: Timothy Wong
Scikit-bio offers a great toolkit for working with biological data that has a tree-like structure. Although the current implementation, TreeNode, can accommodate a large variety of use-cases, its pointer-based nature limits its scalability to large trees with billions or even trillions of nodes. This project aims to remedy this problem by implementing a succinct data structure for trees based on its balanced parentheses representation, such that the memory usage and the time required for operations are minimized. There are two main components to this project. The first component pertains to the implementation of the data structure proposed in using Numba and NumPy based on improved-octo-waddle, a previous implementation using Cython. The other component relates to integrating the new data structure into the back-end of the existing TreeNode class within the library to enable compatibility with the existing APIs. The core deliverables include: an implementation of the new data structure that can be interfaced via TreeNode, an extensive suite of tests for correctness and benchmarking to ensure performance uplifts, and documentation to guide users on the usage of the new data structure. As a potential extension, GPU support will be investigated along with the exploration of an alternative solution proposed by Córdova, J. and Navarro, G. (2016).
Mentors: Student: Mudit Maheshwari
The current gate-level circuit simulator in qutip-qip utilizes a full sparse Qobj representation of each gate, expanding local gates to the Hilbert space before applying them to the state vector via matrix multiplication. This approach encounters significant memory and computational bottlenecks as the number of qubits increases. This project aims to address that limitation by refactoring the circuit simulator and adding tensor based executing via the recently added qutip.einsum. PR #225 in qutip-qip has made efforts to introduce a foundational tensor contraction approach, but it is only limited to the numpy backend and lacks support for GPU acceleration. Hence, this project aims to further implement JAX integration to the backend using qutip.einsum with JIT compilation for speedups on both CPU and GPU, As well as,to implement GPU accelerated circuit simulation using qutip-cuquantum.
Mentors: Student: Sumaiya Islam
Panel and HoloViews are powerful Python visualization libraries but they have no MCP integration today. AI agents can write visualization code but the result never renders inline in the chat. There is no way to create, update, or interact with a live dashboard from within an LLM conversation. This project builds an MCP Apps-compatible server for Panel and HoloViews that changes that. MCP Apps let tools return sandboxed, embedded interfaces that render directly inside MCP clients like VS Code, Claude Desktop, and Cursor, with bidirectional communication between the interface and the host. The goal is to turn Panel and HoloViews into first-class interactive UIs inside any agent conversation. The project will deliver three things: a CSP-compliant sandboxed rendering path for MCP Apps covering the full hvPlot and HoloViews chart surface, a bidirectional postMessage protocol so widgets can communicate with the agent in real time, and cloud-native dataset support for Parquet, Zarr, and the Earthmover marketplace. All work will be delivered as tested, documented PRs to the HoloViz ecosystem.
Mentors: Student: Jocho-Smith
This project redesigns sbi's potential function API to eliminate stateful behavior, inconsistent IID handling, and device management bugs. The new architecture introduces a three-layer design: (1) stateless core protocol using typing.Protocol for composability, (2) explicit observation binding utilities with clear IID semantics, and (3) high-level convenience factories. This enables seamless integration with NumPyro, PyMC, and PyTorch samplers while maintaining backward compatibility through a phased migration strategy. Deliverables include refactored potential classes, comprehensive tests, migration guide, UML diagrams, and tutorial notebooks demonstrating ecosystem integration. The redesign fundamentally improves maintainability, extensibility, and reliability for both novice users and advanced practitioners.
Mentors: Student: eclipse
Standard Bayesian posteriors degrade under model misspecification by collapsing to point estimates. This project implements Predictively Oriented (PrO) posteriors in PyMC, a theoretically principled sampler that provides uncertainty calibrated to predictive reality rather than parameter estimation. The implementation includes a scoring-rule abstraction, Wasserstein gradient computation via PyTensor, vectorised particle simulation, and ArviZ integration with tutorials demonstrating robustness on real datasets.