Project Portfolio
automate intelligent, context-aware code improvements beyond syntax checking. analyzes abstract semantics to suggest and execute complex transformations — renaming, restructuring, simplification — vastly improving code quality and maintainability.
- Frontend: React interface for inputting code, selecting refactoring types, and reviewing changes.
- Backend: FastAPI service handling concurrent requests with PostgreSQL for data persistence.
- AI Core: Gemini API with advanced prompt engineering for context-aware pattern detection and semantic transformation.
improved accuracy and reliability of cancer grading by integrating multiple data modalities, grounded in the principles of the "IHC Matters" research paper, developed under MMVL Lab.
- Model: Attention-based Multiple Instance Learning (MIL) for whole-slide imaging data.
- Innovation: Two-stage pipeline with feature fusion and generative models for enhanced representation learning.
- Scale: Trained and validated on a dataset of over 5,000 medical images.
high-fidelity, from-scratch implementation of the Transformer-based Set Prediction Network as a foundational and flexible module for set prediction in research-focused deep learning pipelines.
- Core: Built from fundamental PyTorch components with no high-level wrappers, ensuring maximum control over layer implementation, optimization, and debugging.
- Application: Designed as a replaceable module within complex research architectures with emphasis on clear structure and performance on scientific datasets.
high-performance, from-scratch text editor in Rust focusing on low-level systems programming principles — efficiency, safety, extensibility. first major Rust project; AI assistance used for code generation and debugging.
- Core: Entire editor built from fundamental Rust components, prioritizing low-level control with no external runtime dependencies.
- Application: Extensible editor suitable for programming and writing tasks with robust terminal rendering via crossterm.
diffusion model built from scratch in PyTorch and trained on CIFAR, exploring core generative modeling principles at the implementation level.
- Core: Full diffusion pipeline from fundamental components — forward process, noise schedule, U-Net denoiser — no high-level wrappers.
- Application: Serves as a flexible base for experimenting with image generation and score-based generative models.
basic ZK-SNARK protocol implemented in Rust using ArkWorks, primarily as a Rust learning exercise in cryptographic systems. AI assistance used for code generation and debugging.
- Core: End-to-end ZK-SNARK circuit setup, proving, and verification using the ArkWorks library.
- Application: Usable as a base for more complex cryptographic tasks and zero-knowledge applications.