Projects

### Overview: Hypster is a lightweight configuration framework designed specifically for optimizing AI and ML workflows. It simplifies the process of managing and optimizing AI configurations with a Pythonic API, hierarchical and conditional configurations, type safety, and built-in hyperparameter optimization. ### Key Features: - **Pythonic API**: Intuitive and minimal syntax that feels natural to Python developers. - **Hierarchical, Conditional Configurations**: Support for nested and swappable configurations. - **Type Safety**: Built-in type hints and validation. - **Hyperparameter Optimization Built-In**: Native, first-class optuna support. - **Rules As Config**: Define replayable WHEN/THEN rule lists with built-in notebook controls. ### Tech Stack: - Python ### Why It Matters: Hypster simplifies the configuration process for AI and ML workflows, making it easier for developers to manage and optimize their models. Its Pythonic API, hierarchical configurations, and built-in optimization features make it a valuable tool for anyone working with AI and ML projects.

### Overview: Orion is a deep-learning-based system designed for automated detection and classification of military vehicles in video data. It integrates visual recognition, motion analysis, and tracking modules to provide real-time situational awareness in complex environments. ### Key features: - **Automated Vehicle Detection and Classification**: Orion can identify and classify military vehicles in video data. - **Real-Time Tracking**: Tracks military vehicles in videos in real-time. - **Custom Dataset**: Fine-tuned YOLO12 model for object detection on a custom dataset of military vehicles. - **Models**: Provides YOLO12 models fine-tuned on a custom dataset of military vehicles with 4 classes. ### Tech stack: Python ### Why it matters: Orion offers a powerful tool for automating the recognition and tracking of military vehicles in video data, enhancing situational awareness in complex environments.

### Overview: SelfJev is a self-hosted decision model that turns text and questions into structured decisions using a 4B model running on your own infrastructure. It provides a Python SDK and a Jev-compatible HTTP API, making it easy to integrate into applications. ### Key Features: - **Self-hosted Decision Model**: Runs on your own GPU, providing complete control over the model and infrastructure. - **Python SDK**: Simplifies integration into applications. - **Jev-Compatible API**: Compatible with the Jev decision model, allowing seamless integration with existing systems. ### Tech Stack: Python ### Why It Matters: SelfJev enables developers to turn text and questions into structured decisions, making it easier to integrate into applications and provide contextually relevant answers directly from your infrastructure.

Overview: This repository provides a seamless integration between LangChain and Milvus, enabling efficient vector storage and retrieval for AI applications like semantic search, recommendation systems, and RAG. Key features: - Vector Storage - Similarity Search - Hybrid Search - Maximal Marginal Relevance - Multiple Vector Fields - Sparse Embeddings - Built-in Functions (BM25) - Async Support Tech stack: Python Why it matters: This integration simplifies the development of AI applications by providing a powerful vector database for efficient vector search and retrieval, making it easier to build sophisticated semantic search, recommendation, and RAG systems.

### Overview: The project aims to classify and localize affected areas in Chest Radiographs (X-Rays) using Deep Learning techniques. It utilizes the COVID-19 Radiography Database, which contains 21,165 Chest X-Rays belonging to 4 different classes: COVID-19, Lung Opacity, Normal, and Viral Pneumonia. The project fine-tunes three top-scoring CNN architectures (VGG-16, ResNet-18, and DenseNet-121) on the dataset and employs Gradient-weighted Class Activation Maps (Grad-CAM) to localize affected areas. ### Key Features: - Fine-tuning of three CNN architectures (VGG-16, ResNet-18, DenseNet-121) on the COVID-19 Radiography Database. - Evaluation and comparison of results obtained from different architectures. - Localization of affected areas in Chest X-Rays using Grad-CAM. - Dataset exploration, split, and evaluation of results. ### Tech Stack: - PyTorch ### Why it Matters: The project aims to bridge the gap in diagnosing COVID-19 from other viral pneumonia cases, making it easier for radiologists to recognize the disease. By utilizing Deep Learning techniques, the project seeks to improve the accuracy and reliability of diagnosing COVID-19 through Chest X-Rays.

Overview: DeepSpot-M is a multimodal foundation model designed to predict transcriptome-wide virtual spatial transcriptomics from histology images. Key features: - Maps a 224x224 H&E tile to spatial gene expression. - Uses a LoRA-adapted pathology foundation backbone and a gene decoder for cross-attention. - Generates gene-specific output projections from biological embeddings. - Enables zero-shot prediction of genes at inference time. Tech stack: Python, PyTorch, Hugging Face Transformers. Why it matters: DeepSpot-M provides a powerful tool for generating virtual spatial transcriptomics data from histology images, making it accessible without requiring specialized hardware.

### Overview: Spokestack: A React Native Plugin for Voice Interface Development Spokestack is a plugin for React Native that integrates voice recognition, wake-word detection, and natural language understanding capabilities. This plugin is designed to enhance the user experience by providing a voice interface for your React Native app. Whether you're developing a mobile app or a web application, Spokestack offers a seamless way to integrate voice functionality. ### Key Features: - **Speech Recognition**: Translates spoken commands into text. - **Wakeword Detection**: Recognizes voice commands as wake words. - **Natural Language Understanding**: Analyzes and interprets spoken language for context and intent. - **Text-to-Speech Synthesis**: Converts text into spoken words. ### Tech Stack: TypeScript Spokestack is built with TypeScript, a modern and statically typed programming language that supports type checking at compile time. This ensures type safety and helps catch errors early in the development process. ### Why It Matters: Enhancing User Experience Spokestack's integration with React Native makes it an ideal choice for developers looking to add voice functionality to their applications. Whether you're developing a mobile app, a web-based application, or a hybrid app, Spokestack provides a powerful tool for enhancing the user experience by allowing users to interact with your app through voice commands. ### Repository: spokestack/react-native-spokestack The repository for Spokestack is hosted on GitHub, where it has received 60 GitHub stars. This indicates that the plugin has been widely adopted and is well-received by the community. ### Requirements

Overview: KMP-Stable-Diffusion is a Kotlin-based multiplatform application that generates images using Stability AI. This project is designed to be a powerful tool for developers looking to integrate image generation into their applications. Key features: - Text-to-image and image-to-image functionalities - Support for various styles and modes - Local save functionality - Dark and light mode options Tech stack: Kotlin Why it matters: KMP-Stable-Diffusion provides a user-friendly interface for developers to easily integrate image generation into their applications, making it a valuable tool for creating dynamic and visually appealing projects.

### Overview: This repository serves as an endpoint server for efficiently serving quantized open-source LLMs for code, integrating support for quantized models into the `llm-vscode` extension. It is designed to be more accessible and cost-effective, even on smaller GPUs and CPUs, by leveraging the `llm-vscode` extension. ### Key features: - Integrates support for quantized open-source LLMs for coding tasks. - Provides a self-hosted code completion service. - Runs on `localhost` by default on port `8000`. - Supports command line arguments for flexibility. ### Tech stack: Python ### Why it matters: This solution offers a more accessible and cost-effective alternative to existing code completion services, making it ideal for developers who need to quickly generate code snippets and comments.

Overview: AI Lawyer is an intelligent reasoning legal assistant powered by DeepSeek R1 and Ollama RAG, designed to streamline legal research and document analysis. Key features: - Upload and analyze legal documents (PDFs) - Retrieve relevant legal information using FAISS vector database - Answer legal questions using DeepSeek R1 with Groq - Summarize legal documents - Generate downloadable AI-generated legal reports Tech stack: Python Why it matters: AI Lawyer leverages advanced reasoning models and retrieval-augmented generation to provide precise legal insights and structured document analysis, enhancing the reliability and accuracy of legal research and document processing. Repository: AbhaySingh71/AI-Lawyer-RAG-with-Deepseek Primary language: Python GitHub stars: 59 GitHub description: AI Lawyer is an intelligent reasoning legal assistant powered by DeepSeek R1 and Ollama RAG, designed to streamline legal research and document analysis. By leveraging retrieval-augmented generation (RAG), it provides precise legal insights, and contract summarization. With an intuitive Streamlit-based UI, analyze legal documents.

Overview: Vico is a unified solution for compositional video generation that ensures all concepts are represented properly by equalizing the information flow of text tokens. Key features: - Compositional Video Generation as Flow Equalization - Supports multiple diffusion-based video models for compositional T2V and video editing Tech stack: Python Why it matters: Vico significantly enhances the compositional richness and accuracy of generated videos by preventing any single concept from dominating, ensuring proper representation of input tokens.

Overview: This repository contains an official implementation of the paper "Towards Flexible Multi-modal Document Models" presented at CVPR 2023. It includes pre-trained models, datasets, and training scripts for various experiments, with a focus on flexible multi-modal document models. Key features: - Pre-trained models for Crello and Rico datasets - Training scripts for Crello and Rico datasets - Evaluation scripts for both datasets - Tensorboard monitoring for training progress Tech stack: Python, CUDA 11.3, Tensorflow 2.8 Why it matters: This repository provides a comprehensive solution for researchers and developers interested in flexible multi-modal document models, offering practical tools for both pre-trained model usage and model training.

### Overview: This repository contains the official code for the paper "Stochastic Segmentation with Conditional Categorical Diffusion Models" (ICCV 2023), which presents a novel conditional categorical diffusion model (CCDM) for semantic segmentation. The model addresses the challenge of generating multiple segmentation label maps to account for aleatoric uncertainty in safety-critical domains like medical diagnostics and autonomous driving. ### Key features: - **Conditional Categorical Diffusion Model (CCDM)**: A conditional categorical diffusion model designed for semantic segmentation. - **Training on LIDC Dataset**: The code includes training scripts for semantic segmentation using multiple annotations. - **Evaluation on LIDC Dataset**: Evaluation scripts for the LIDC dataset, including training and evaluation parameters. - **Cityscapes Dataset Support**: The repository supports experiments on the Cityscapes dataset, including a pretrained model and parameter files. ### Tech stack: Python ### Why it matters: The CCDM model proposed in this repository addresses the need for generating multiple segmentation label maps in safety-critical domains, where the true distribution of annotation maps is typically multimodal and high-dimensional. This is particularly relevant in medical diagnostics and autonomous driving, where a single segmentation output may not be sufficient to capture the true uncertainty in the data. The CCDM model aims to predict conditional distributions of labels given the input image, enabling it to generate multiple segmentation maps that account for aleatoric uncertainty, thus improving the robustness and reliability of semantic segmentation models.

Overview: The DeepFakeFace repository provides a curated collection of artificial celebrity faces using cutting-edge diffusion models, addressing the challenge of deepfakes in today's digital landscape. Key features: - DeepFakeFace dataset - Installation instructions for Diffusers, Pytorch, and InsightFace - Generation of deepfake images using Stable Diffusion V1.5, Stable Diffusion Inpainting, and InsightFace - Evaluation of deepfake detection methods using RECCE and DeeperForensics - Performance metrics for deepfake detection accuracy, AUC, and EER - Citing the research paper Tech stack: Python Why it matters: This repository offers a valuable resource for researchers and developers working on deepfake detection, providing a comprehensive dataset and tools for evaluating and improving detection methods.

Overview: SynthGenAI is a package designed to generate synthetic datasets using Large Language Models (LLMs) from various API providers. It aims to address the challenges of data scarcity and high costs by offering a simple and modular solution. Key features: - Supports multiple LLM API providers - Easy to integrate with different datasets - Provides a user-friendly command-line interface - Can be extended with additional API providers Tech stack: Python Why it matters: By leveraging synthetic data, SynthGenAI helps enhance the performance of AI models across various use cases, including pre-training and post-training of Small Language Models (SLMs), while balancing resource efficiency and model effectiveness.

Overview: ProLearn introduces a novel prototype-driven learning framework that addresses the limitations of textual reliance in medical language-guided segmentation, enabling image-only, image-text data mix training and inference with minimal text input. Key features: - Image-only, image-text data mix training - Inference with limited or no textual input - Robust and localized segmentation maps - PSA module preserving attention saliency and lesion coherence - Prototype zoo for easy experimentation Tech stack: Python Why it matters: ProLearn provides a significant improvement over previous work, maintaining performance even under extreme text scarcity, making it highly relevant for real-world applications in medical image analysis.

Overview: 这个项目旨在为DeepSeek提供“眼睛和耳朵”的功能，通过集成多种视觉和语音处理工具，使DeepSeek具备看图、OCR、物体检测、视频理解、语音转写以及分析音乐和读取文档的能力。这个插件通过一键安装即可使用，无需额外配置，只需提供一个API密钥即可调用云端视觉大模型。 Key features: - 看图 - OCR - 物体检测 - 视频理解 - 语音转写 - 分析音乐 - 读取PDF/文档 Tech stack: PowerShell Why it matters: 这个插件为DeepSeek提供了强大的视觉和语音处理能力，使其能够处理各种图像、视频和文档内容，从而提升其在信息检索和理解方面的表现。通过一键安装，用户无需额外配置即可使用，极大地提高了DeepSeek的实用性和易用性。 Repository: Yts1919/dsh-vision-complete GitHub stars: 57 GitHub description: 给DeepSeek补上“眼睛和耳朵”的多模态视觉插件：看图/OCR/物体检测/视频理解/语音转写/截图直读，一键安装（DSH插件）。

### Overview: Seedance 1.0 Pro Prompts: 8 Tested Recipes With Outputs Dreamina — Seedance 1.0 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-1.0-pro`) and the cost per image. This repository provides a comprehensive guide to crafting effective prompts for Seedance 1.0 Pro, including four recipes with detailed instructions and example renders. ### Key Features: - **Model ID:** `seedance-1.0-pro` - **Example Renders:** Four recipes with detailed prompts and example renders. - **Cost per Image:** $0.0184 per second. ### Tech Stack: Python ### Why It Matters: Seedance 1.0 Pro is a powerful tool for image generation, offering a wide range of options and flexibility. This repository provides practical tips and examples to help developers create effective prompts, ensuring high-quality outputs at a reasonable cost. ### Repository: a7m-7rm3vb/doubao-seedance-1-0-pro-prompts-pt ### Primary Language: Python ### GitHub Stars: 43 ### GitHub Description: Dreamina — Seedance 1.0 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-1.0-pro`) and the cost per image. This repository provides a comprehensive guide to crafting effective prompts for Seedance 1.0 Pro, including four recipes with detailed instructions and example renders. ### Primary Language: Python ### GitHub Stars:

### Overview: xKV is a cross-layer Singular Vector Decomposition (SVD) method designed for KV-cache compression. It aims to achieve up to 8x KV-cache compression while maintaining accuracy on long-context tasks. The project was accepted at ICML 2026 and is part of the Cornell University and National Yang Ming Chiao Tung University research team. ### Key features: - **Cross-Layer SVD Compression**: xKV decomposes KV-cache into a shared low-rank subspace across layers. - **Selective Reconstruction (SR)**: xKV-SR enhances performance by selectively reconstructing keys and values during decoding. - **Efficiency Benchmarks**: Released benchmarks for decode latency, end-to-end throughput, and SVD overhead. ### Tech stack: - Python ### Why it matters: xKV offers a simple yet effective solution for KV-cache compression, significantly improving performance without compromising accuracy on long-context tasks.

### Overview: CortexaDB is a lightweight, fast, and hard-durable embedded database designed specifically for AI agent memory. It provides a single-file-like experience with native support for vectors, graphs, and temporal search, ensuring your agent never forgets. ### Key Features: - **Single-File Storage**: No server required, making it easy to integrate into any AI agent. - **Native Support**: Vectors, graphs, and temporal search capabilities. - **Hard-Durable Storage**: WAL-backed storage for zero data loss. - **Fast Ingestion**: 100x faster than traditional methods. - **Hybrid Retrieval**: Search by semantic similarity, structural relationship, and time-based recency. - **Fluent API**: Chainable QueryBuilder for expressive searching and collection scoping. - **Ultra-Fast Indexing**: Uses HNSW for sub-millisecond approximate nearest neighbor search. - **Privacy First**: Completely local, ensuring your agent's memory stays on your machine. ### Tech Stack: Rust ### Why It Matters: CortexaDB provides a middle ground for AI agents, offering a hard-durable, embedded memory engine that runs inside your agent's process. It ensures your agent never forgets, starting instantly with zero overhead, and maintaining millisecond query latencies even as it learns thousands of new facts. ### Installation: - **Python**: `pip install cortexadb` - **Rust**: `cargo add cortexadb` ### Core Capabilities: - **100x Faster Ingestion**: New batch insertion system allows processing 5,