Projects

Overview: This project provides an official JavaScript/TypeScript SDK for the WaveSpeed inference platform, enabling developers to easily integrate image, video, audio, and 3D generation capabilities via the Wavespeed AI API. Key features: - WaveSpeed JavaScript/TypeScript SDK - Supports various models including z-image, z-video, z-audio, and z-3d - Provides a unified and high-performance API - Supports authentication via environment variables or direct key - Offers options for timeout, poll interval, and enable sync mode - Supports upload of images, videos, and audio files - Allows setting client-level retry configuration - Handles request timeouts for individual HTTP requests - Provides task ID and debug information for logging and debugging Tech stack: TypeScript Why it matters: This SDK simplifies the integration of advanced AI capabilities into JavaScript/TypeScript applications, making it easier for developers to leverage WaveSpeed's cutting-edge technology without the need for extensive backend development.

### Overview: This project explores the effectiveness of fine-tuning on specialized datasets to mitigate bias in large language models like Bloomz and FLAN-T5. The objective is to evaluate if fine-tuning on the CrowSPairs dataset can improve bias reduction in these models, which are more recent and larger than BERT and GPT-2. ### Key Features: - Fine-tuning on specialized datasets (CrowSPairs) to reduce bias in large language models. - Comparison of models: FLAN-T5, Bloomz, and others. - Use of HuggingFace metrics for toxicity, regard, and honest metrics. - Presentation of results from pre and post-fine-tuning evaluations. ### Tech Stack: Jupyter Notebook ### Why it Matters: The project aims to provide a proof of concept for bias mitigation techniques in large language models, contributing to the field of AI ethics and fairness.

### Overview: This UEFI DXE driver enables 4G decoding, Resizable BAR, and MMIO64 placement on a 990fx chipset. It is designed to work with a specific BIOS and chipset configuration, ensuring compatibility with the Gigabyte GA-990FX-Gaming Rev 1.0 with AMD SR5690 + SP5100, AMI Aptio IV, and AMD FX (Family 15h) CPUs. ### Key Features: - **4G Decoding**: The module supports decoding of 4G GPUs, enabling them to run at full performance. - **Resizable BAR**: It allows for the allocation of large memory regions, up to 16GB, without the need for kernel patches. - **MMIO64 Placement**: The module places the MMIO64 regions at specific addresses, ensuring optimal performance and compatibility with the BIOS. ### Tech Stack: - **C** - The primary programming language used for the driver development. ### Why It Matters: This project is dedicated to the work of [xCuri0/ReBarUEFI], providing a solution that extends and enhances the capabilities of the original ReBarUEFI driver. The module is designed to work with specific hardware configurations, ensuring compatibility and performance without the need for kernel patches. The validation screenshots and bench results demonstrate the effectiveness of the module, particularly in supporting GPUs designed for servers with terabytes of memory.

### Overview: This repository contains the evaluation system for a paper on using Prefill-Guided Thinking (PGT) to detect AI-generated images with Vision-Language Models (VLMs). ### Key Features: - **Prefill-Guided Thinking (PGT)**: Improves detection accuracy by guiding VLMs with specific prompts. - **Three VLMs**: Tested across 3 VLMs and 16 different image generators. - **Three Datasets**: Evaluated on D3, DF40, and GenImage datasets. ### Tech Stack: - Python 3.11 - PyTorch 2.7.1 - vLLM 0.10.1 - License: MIT ### Why It Matters: - **Improves Detection Accuracy**: PGT improves detection accuracy by up to 24% in Macro F1 without any training or fine-tuning. ### Repository: - Repository: Zoher15/PGT - GitHub Stars: 7 - GitHub Description: Code for Prefilled responses enhance zero-shot detection of AI-generated images ### Citations: - Paper: [arXiv:2506.11031](https://arxiv.org/abs/2506.11031) - Citation: Kachwala et al., "Prefill-Guided Thinking for zero-shot detection of AI-generated images," 2025 ### Quick Start: - **Installation**: See [SETUP.md](SETUP.md) for complete

### Overview: Lao is a headless engine and CLI designed for chaining local plugins into DAG-based workflows. It supports dynamically loaded shared libraries built against a stable C ABI, parallel execution, and fully offline operation. Plugins can be validated, parsed, and executed through a YAML workflow, with features like retries, caching, and loops. ### Key Features: - **Dynamic Loading:** Plugins are loaded using `dlopen` from shared libraries (`.so`, `.dylib`, `.dll`) that export a known C vtable. - **Workflow Parsing:** The CLI parses a YAML workflow into a Directed Acyclic Graph (DAG) and validates it for plugin availability and I/O type compatibility. - **Parallel Execution:** Steps can be executed in parallel based on their dependency levels. - **Retry Mechanism:** Steps can be retried with backoff for failed executions. - **Output Caching:** Step outputs can be cached on disk for efficient reuse. - **Conditions and Loops:** Supports conditions and loops for more complex workflows. ### Tech Stack: Rust ### Why it Matters: Lao simplifies the process of creating and managing workflows, making it easier to integrate local plugins into complex, offline workflows. Its dynamic loading and parallel execution capabilities make it suitable for a wide range of applications, from audio transcription to natural language processing tasks.

### Overview: Corpus2GPT is a project designed to enable users to train their own GPT models on diverse datasets, including local languages and various corpus types. It leverages Keras and is compatible with TensorFlow, PyTorch, or JAX backends, making it accessible for users of different skill levels. The project is committed to innovation in language model research, offering a user-friendly platform and comprehensive documentation. ### Key Features: - **Modular Design**: Corpus2GPT is modular, making it easy to navigate, modify, and understand. - **Versatile Datasets**: Supports training on diverse datasets including local languages and various corpus types. - **Compatibility**: Compatible with Keras, TensorFlow, PyTorch, and JAX backends, enabling flexibility and benchmarking. - **Documentation**: Comprehensive documentation and support for various language corpora, backends, and scaling solutions. - **Ease of Use**: Simple interface and intuitive user experience, making it accessible to users of all backgrounds. ### Tech Stack: - **Keras**: The primary technology stack used for training and evaluating models. ### Why It Matters: Corpus2GPT is a pioneering project that revolutionizes language model research by offering a user-friendly platform. It prioritizes accessibility and ease of use, making it ideal for a diverse range of users, from NLP researchers and data scientists to language enthusiasts and industry professionals. With its modular design and compatibility with multiple backends, Corpus2GPT sets a new standard in language model research and development.

Overview: This guide provides a straightforward way to try out the Vector Search feature in Azure Cosmos DB for NoSQL, using a simple dataset of movies to find similar movies based on a given criteria. The app is implemented in Python, TypeScript, .NET, and Java, with instructions on setting up, loading data, and executing similarity search queries. Key features: - Vector Search feature implementation - Simple dataset of movies - Similarity search based on vector embeddings - Implemented in Python, TypeScript, .NET, and Java - Instructions for setting up, loading data, and executing queries Tech stack: Java Why it matters: This guide is useful for developers looking to quickly and easily try out the Vector Search feature in Azure Cosmos DB for NoSQL, providing a practical example of how to implement and use the feature in various programming languages.

### Overview: NInfer-3080 is a 27B model optimized for a single 10 GB RTX 3080 GPU, achieving 139 tok/s on short answers and 94 tok/s with 61K tokens in context, running on native Windows with MSVC and CUDA 13.8 or 13.x. ### Key features: - **Fastest Performance**: 139 tok/s on short answers, 94 tok/s with 61K tokens in context (MTP, 64K window). - **Longest Window**: 96K window that still decodes at 53 tok/s when nearly full. - **Prompt Filling**: 1,400 to 1,900 tok/s prefill. - **Runs**: natively on Windows with MSVC and CUDA 13.8 or 13.x; no WSL. ### Tech stack: C++ ### Why it matters: NInfer-3080 is optimized for a specific GPU and Windows environment, providing a significant performance boost for inference tasks on a single RTX 3080 card.

Overview: This repository contains code for analyzing the representations produced by contextual language models from a topological perspective, focusing on changes in the local intrinsic dimension (LID) of the model's hidden states during training and fine-tuning. Key features: - Studying the local intrinsic dimension (LID) of contextual language models' hidden states - Predicting when the model's training capabilities are exhausted through mean local dimension - Demonstrating practical heuristics for informed decisions in model configuration Tech stack: Python Why it matters: This approach provides insights into the training dynamics and generalization ability of language models, offering practical guidance for practitioners in configuring models for specific applications.

### Overview: This repository provides a template for building microservice-based applications using a frontend, backend, LLM serving engine (e.g., vLLM), and nginx as a reverse proxy. It is designed to help developers quickly set up and deploy applications that utilize large language models (LLMs) alongside traditional web services. ### Key Features: - **Frontend**: Responsible for the frontend part of the application. - **Backend**: Handles the business logic and communicates with the LLM server. - **LLM Server**: A serving engine (such as vLLM) that provides API endpoints for large language model inference. - **Nginx**: Acts as a reverse proxy to route traffic between the frontend and backend. ### Tech Stack: - HTML ### Why it Matters: This template simplifies the process of building microservices applications, making it easier to integrate LLMs into traditional web services. It provides a clear structure and dependencies, ensuring developers can focus on application logic without worrying about the underlying infrastructure.

Overview: This project compares two BERT-based models, Spanish BERT and Multilingual BERT, for sentiment analysis on Amazon reviews in Spanish. Key features: - Sentiment analysis on Amazon reviews in Spanish - Comparison of Spanish BERT and Multilingual BERT models - Fine-tuned BERT models for sentiment classification Tech stack: Jupyter Notebook Why it matters: The Spanish BERT model demonstrates superior performance in sentiment analysis on Amazon reviews in Spanish, highlighting the potential of multilingual BERT models in non-English language tasks.

### Overview: This GitHub Action automates the deployment of machine learning models and inference scripts to Algorithmia, streamlining the continuous integration and deployment process. ### Key Features: - Automatically uploads and updates ML models and inference scripts to Algorithmia. - Supports both Jupyter notebook training and model file upload workflows. - Integrates with Algorithmia's scalable inference endpoint, enabling model inference. - Manages metadata in `model_manifest.json` for model versioning and verification. ### Tech Stack: Python ### Why it Matters: This action simplifies the deployment process, ensuring models are always up-to-date and easily accessible, while maintaining version control and metadata integrity.

Overview: This project is a production-grade genpark-beam-search-autoregressive-decoder-skill skill designed for AI agents, focusing on enhancing the efficiency and accuracy of language generation tasks. Key features: - Auto-regressive language generation - Beam search optimization - Fine-tuning for specific use cases Tech stack: Python Why it matters: This skill is crucial for developers looking to improve the performance and reliability of AI-driven language models, making it an essential tool for applications requiring high-quality text generation.

Overview: This project optimizes the prompt prefix cache by matching longest common prompt token prefixes, reducing redundant prefill computation and minimizing TTFT (Time to First Token). Key features: - Optimizes prompt prefix cache - Matches longest common prompt token prefixes - Reduces redundant prefill computation - Minimizes TTFT Tech stack: Python Why it matters: This project helps developers reduce redundant computations and improve the efficiency of their prompt prefix cache, leading to faster and more optimized performance.

Overview: This project introduces a production-grade hierarchical navigable small-world skill for AI agents, optimized for efficiency and scalability. Key features: - Hierarchical indexing - Small-world graph structure - Efficient data retrieval Tech stack: Python Why it matters: This skill significantly enhances the performance and scalability of AI agents, making it essential for advanced applications requiring fast and accurate data retrieval.

Overview: This project introduces a novel vector engine called IVF-Flat, which is designed to efficiently index and search large datasets using inverted file indexing and flat quantization techniques. The engine partitions the dataset into Voronoi cells and ranks the top-K Euclidean nearest neighbors in a highly optimized manner. Key features: - Efficient inverted file indexing - Flat quantization for compact storage - Fast top-K Euclidean ranking Tech stack: Python Why it matters: This project offers a significant improvement in vector indexing and search performance, making it particularly useful for applications requiring high-speed and accurate nearest neighbor retrieval in large datasets.

Overview: This project introduces a virtual paged attention key-value cache memory manager, designed to efficiently manage and allocate memory for large datasets. It includes features such as logical block mapping, non-contiguous physical page allocation, and zero-copy fragmentation tracking to optimize performance and reduce memory usage. Key features: - Logical block mapping - Non-contiguous physical page allocation - Zero-copy fragmentation tracking Tech stack: Python Why it matters: This project is crucial for developers looking to optimize memory management in their applications, particularly for handling large datasets efficiently and reducing memory fragmentation.

### Overview: This repository contains the code and benchmark datasets for the paper "Token Alignment via Character Matching for Subword Completion" presented at ACL Findings 2024. The paper proposes a novel technique called token alignment to mitigate issues arising from tokenization artifacts, particularly relevant for applications like code completions where LLMs often generate sub-optimal responses due to ending partial tokens. ### Key Features: - Token Alignment: A technique to align tokenized inputs to their original form, improving the quality of generated responses. - Character Trie Optimization: Pre-built character trie and mask cache to optimize inference efficiency. - Extensive Experiments: Conducted on code completions and general NL tasks to validate the improvements of the proposed token alignment. ### Tech Stack: - Python ### Why It Matters: The paper addresses a significant issue in the use of large language models (LLMs) in practical applications, such as code completions, where the model's output can be suboptimal due to tokenization artifacts. By proposing token alignment, the authors aim to enhance the quality of generated responses, leading to better user experiences and more effective applications of LLMs. ### Repository: - **Repository**: amazon-science/tokenalign - **Primary Language**: Python - **GitHub Stars**: 7 - **GitHub Description**: Token Alignment via Character Matching for Subword Completion (ACL Findings 2024) - **Evaluation of Public Models**: Code to evaluate public models on MBXP datasets. - **Usage Example**: ```python CUDA_VISIBLE_DEVICES=0,1

### Overview: This repository contains the fine-tuned and compared three multilingual language translation models: MBart, AI4Bharat, and Helsinki-NLP. The models were trained on the English to Marathi and Marathi to English translation tasks. ### Key Features: - Fine-tuned MBart, AI4Bharat, and Helsinki-NLP models for English to Marathi and Marathi to English translation. - Leveraged Hugging Face's transformers library for model creation and fine-tuning. - Achieved remarkable results with Helsinki-NLP, which had a loss rate of 0.5174, surpassing other models. ### Tech Stack: - Python - Hugging Face Transformers Library ### Why it Matters: This project showcases the effectiveness of fine-tuning multilingual language models on specific translation tasks, demonstrating the potential for improving language translation accuracy across different languages.

Overview: This project is a Spring Boot-based implementation of a Retrieval-Augmented Generation (RAG) system using Spring AI 2.0.2, PGVector, and OpenAI. It includes conversation memory, tool calling, streaming capabilities, and an Open WebUI for integration with OpenAI-compatible APIs. Key features: - Spring Boot 4.1.2 with Spring AI 2.0.2 - Java 25 (virtual threads enabled) - PostgreSQL 17 + PGVector 0.8.7 - Open WebUI 0.11.4 (optional chat interface) Tech stack: Java Why it matters: This project showcases advanced AI integration with Spring Boot, leveraging PGVector for efficient vector-based queries and OpenAI for seamless integration with Open WebUI, making it suitable for developers looking to implement AI-driven conversational applications.