Projects

Overview: Flow Agent is a local server and command-line interface (CLI) toolkit designed to generate high-quality AI images and videos using Google Flow. Key features: - Text → Image: Supports aspect ratios 1:1, 16:9, 9:16, 4:3, 3:4, and square, generating up to 4 variations at once. - Text → Video: Generates videos with durations of 4s, 6s, 8s, and 10s, with quality options for 360p and 720p. - Image → Video: Animates any image into a video clip. - Multi-Account: Allows adding multiple Google accounts to pool credits. - OpenAI Compatible: A drop-in replacement for OpenAI's API, compatible with the same base URL. - History & Stats: Logs every generation to local SQLite with credit tracking. - Self-Diagnostics: Provides `./bin/flow doctor` to diagnose and fix issues. Tech stack: JavaScript, Python, FastAPI, MCP, SQLite. Why it matters: Flow Agent simplifies the process of generating high-quality AI images and videos from Google Flow, making it easy to integrate into various applications and workflows.

### Overview: The RT-X repository is a PyTorch implementation of the models RT-1-X and RT-2-X from the paper "Open X-Embodiment: Robotic Learning Datasets and RT-X Models". This project aims to provide a practical tool for robotic learning datasets and RT-X models, focusing on the integration of PyTorch for efficient and accessible model training and evaluation. ### Key Features: - **RTX-1**: Implements a SOTA transformer architecture, using a Film-EfficientNet backbone with Film layers inserted. - **RTX-2**: Outputs text tokens from images and text inputs, suitable for applications requiring multi-modal text generation. - **EfficientNetFilm**: Extracts features from images using a pre-trained EfficientNet model. ### Tech Stack: The repository uses Python with PyTorch for the implementation and testing. ### Why It Matters: This project is significant for researchers and developers interested in robotic learning datasets and RT-X models, providing a practical tool for implementing and testing these models in a Python environment. ### Repository: kyegomez/RT-X - **GitHub Stars**: 246 - **Primary Language**: Python - **GitHub Description**: Pytorch implementation of the models RT-1-X and RT-2-X from the paper: "Open X-Embodiment: Robotic Learning Datasets and RT-X Models" - **GitHub License**: MIT ### Installation: ```bash pip install rtx-torch ``` ### Usage: - **RTX1**: Implements a SOTA transformer architecture using a

Overview: SatoriDB is an embedded, high-performance vector database designed for handling billion-scale datasets with exceptional recall and predictable latencies. Key features: - Embedded architecture - Two-tier search system: hot index in RAM, cold vector data on disk - Automatic clustering of vectors - CPU-pinned Glommio workers for parallel scanning - SIMD acceleration for efficient distance computations - Configurable durability options - Persistent storage using Walrus and RocksDB - Linux-only, requires io_uring kernel version 5.8+ Tech stack: Rust Why it matters: SatoriDB excels in ANN search, providing 95%+ recall and predictable latencies for billion-scale datasets, making it ideal for applications requiring fast and accurate nearest neighbor queries. Repository: nubskr/satoriDB GitHub stars: 245 GitHub description: High performance embedded vector database

Overview: This repository contains a PyTorch implementation of the Deep Audio-Visual Speech Recognition model, focusing on the AO, VO, and AV models trained on the LRS2 dataset for speech-to-text transcription. Key features: - PyTorch implementation - Training of three models: AO, VO, and AV - LRS2 dataset for pre-training and testing Tech stack: PyTorch Why it matters: This model provides a practical solution for speech-to-text transcription using audio and video inputs, offering a significant improvement in accuracy over traditional methods. Repository: smeetrs/deep_avsr GitHub stars: 245

Overview: This repository contains the official implementation of CFG++, a method designed to address the limitations of Classifier-Free Guidance (CFG) in diffusion models. The approach aims to correct the off-manifold problem, leading to improved sample quality, better adherence to text, and enhanced invertibility. Key features: - Improved sample quality - Better adherence to text - Smooth, straighter PF-ODE trajectory - Enhanced invertibility Tech stack: Python Why it matters: CFG++ offers a significant improvement over CFG by correcting the off-manifold problem, resulting in better performance in text-to-image generation, DDIM inversion, editing, and solving inverse problems, making it a valuable tool in fields that utilize text guidance. Repository: CFGpp-diffusion/CFGpp GitHub stars: 245

### Overview: dgx-spark-qwen38 is an open-source installer that enables the quick and efficient conversion of NVIDIA DGX Spark or any GB10 machine into a private AI server. This project simplifies the process of setting up a private AI infrastructure, allowing users to deploy models and APIs without the need for complex configurations or manual setup. ### Key Features: - **One Command Install**: Easily turn a DGX Spark or GB10 machine into a private AI server with a single command. - **Model Measured**: Measured performance metrics for various models, including Qwen3.8-27B and Qwen3.8-Flash-Next 176B. - **OpenAI and Anthropic APIs Compatibility**: Supports OpenAI and Anthropic APIs, enabling seamless integration with popular AI tools. - **Local Cockpit**: A user-friendly interface for monitoring and managing the AI server, including a live memory display and a coding agent. - **Pinned and Proven**: Ensures that the AI lanes prove their effectiveness and that the installation process is complete. ### Tech Stack: - Python ### Why It Matters: - **Ease of Use**: The one-command installation process simplifies the setup of a private AI server, making it accessible to users with varying levels of technical expertise. - **Performance**: Measured performance metrics ensure that the AI lanes perform optimally, providing a reliable and efficient setup. - **Compatibility**: Supports OpenAI and Anthropic APIs, making it compatible with a wide range of AI tools and frameworks. - **Pinned and

### Overview: The MambaInLlama repository contains the code and released models for the distillation approach described in the NeurIPS 2024 paper. It aims to distill a large Transformer into a Hybrid Mamba model while preserving generational quality with minimal resources. ### Key features: - **Hybrid Mamba Model**: A distillation approach that combines the strengths of a large Transformer and a smaller model. - **Fast Inference**: Utilizes a lightweight model for fast inference and speculative decoding. - **End-to-End Training**: Minimizes KL divergence loss between the student and teacher models, focusing on the entire model. - **Instruction Tuning**: Optional step to fine-tune instructions for better performance. ### Tech stack: Python ### Why it matters: This repository provides a practical solution for distilling large Transformer models into smaller, more efficient ones, making it ideal for both base models and chat models. The approach is particularly useful for reducing training time and resource requirements, making it accessible even for limited computational resources.

Overview: HYUFA is an AI finance assistant designed for university students and young professionals. It offers a range of features including personalized financial advice, investment analysis, portfolio recommendations, and financial calculations. Key features: - AI Finance Chat - Risk Profile Analysis - Portfolio Recommendation - Finance Calculator Tech stack: HTML, CSS, JavaScript, and possibly some backend technologies for integration with a database or API. Why it matters: HYUFA aims to make financial management accessible and engaging for students, helping them understand basic financial concepts and build healthy financial habits.

Overview: FusionBench is a comprehensive benchmark suite designed to evaluate and compare various deep model fusion techniques. It aims to provide a detailed comparison across different datasets and tasks. Key features: - Comprehensive benchmark suite - Evaluation of model fusion techniques - Detailed performance metrics - Support for diverse datasets and tasks Tech stack: Python Why it matters: FusionBench offers a valuable resource for researchers and developers looking to evaluate and optimize model fusion techniques across various applications and datasets, ensuring that their models perform effectively and efficiently.

Overview: This repository contains Jupyter notebooks designed to support the "From Beginner to LLM Developer" course by Towards AI, focusing on techniques for enhancing AI models with retrieval-based methods. Key features: - Data retrieval techniques - Model integration with retrieval systems - Practical applications of RAG in real-world scenarios Tech stack: Jupyter Notebook Why it matters: This repository provides a comprehensive guide for students and professionals interested in AI and natural language processing, focusing on the practical applications of retrieval-augmented generation techniques.

### Overview: D3PO is a code repository for the paper "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model" presented at CVPR 2024. This repository provides a framework for fine-tuning diffusion models using human feedback directly, eliminating the need for a reward model. ### Key Features: - Fine-tuning diffusion models with human feedback. - Directly fine-tuning the model without the need for a reward model. - Supports both training with and without a reward model. - Utilizes `accelerate` for distributed training. - Provides examples of training with and without a reward model. - Includes sample image generation and human feedback annotation capabilities. ### Tech Stack: Python 3.10 or newer. ### Why It Matters: This repository offers a practical solution for fine-tuning diffusion models through human feedback, making the process more accessible and efficient. It simplifies the traditional reward model-based fine-tuning methods, potentially leading to better model performance and user satisfaction. ### Repository: yk7333/d3po ### GitHub Stars: 244 ### GitHub Description: [CVPR 2024] Code for the paper "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model"

Overview: FE2E is a foundation model designed for monocular dense geometry prediction, specifically for depth and normal estimation. It leverages an advanced image editing model to achieve strong zero-shot performance across both tasks. Key features: - Monocular Depth Estimation - Monocular Normal Estimation Tech stack: Python Why it matters: FE2E addresses the need for efficient and accurate depth and normal estimation in a monocular setting, making it valuable for applications requiring high-quality 3D reconstruction from single images.

### Overview: Mantis is an LLaMA-3-based large multimodal model designed to tackle multi-image vision language tasks. It achieves state-of-the-art performance on five multi-image benchmarks and maintains strong single-image performance comparable to CogVLM and Emu2. ### Key Features: - **Interleaved Text and Image Inputs**: Mantis processes both text and image inputs in an interleaved manner. - **Academic-Level Training**: The model is trained on 36 hours of academic-level resources, specifically on 16xA100-40G. - **State-of-the-Art Performance**: Achieves the best results on NLVR2, Q-Bench, BLINK, MVBench, and Mantis-Eval benchmarks. - **Strong Single-Image Performance**: Maintains strong single-image performance on par with CogVLM and Emu2. ### Tech Stack: Python ### Why It Matters: Mantis represents a significant advancement in the field of multi-image vision language tasks, offering a more efficient and effective approach to training large models on complex multi-image datasets.

Overview: J.A.R.V.I.S. 2.0 is an AI assistant project that utilizes small models and Gemini for hybrid hybridization, providing weather reports, summarizing news, sending emails, and more through voice commands. It supports integration with RAG and efficient memory, and includes features like image generation and phone call automation using ADB. Key features: - Voice Activation - Speech Recognition - AI Responses - Task Execution (email, summarizing weather, data analysis, personalization, news reading, image generation, database functions, phone call automation, AI-based task execution, website & application automation, image processing using Gemini) - Timeout Handling - Automatic Input Processing - Multiple Function Calls - Retrieval-Augmented Generation (RAG) for knowledge-based interactions Tech stack: Python Why it matters: This project showcases the integration of small models and Gemini for hybrid hybridization, providing a versatile AI assistant with features that can handle various tasks through voice commands and efficient memory integration.

Overview: k8ssandra-operator is a Kubernetes operator designed to manage Apache Cassandra and DataStax Enterprise (DSE) on Kubernetes clusters, providing a turnkey solution for deploying and managing multiple data centers across multiple Kubernetes clusters. Key features: - Automated deployment of multiple Apache Cassandra data centers across multiple Kubernetes clusters. - Support for multi-region deployments. - Integration with DataStax Enterprise features such as advanced security, search, and graph. - Support for Change Data Capture (CDC) using Kubernetes manifests. - Monitoring of Apache Cassandra metrics via an agent and integration with Prometheus and Mimir. - Automated repair schedules using Reaper for regular data consistency checks. - Backup and restore capabilities using Cassandra Medusa. Tech stack: Go. Why it matters: k8ssandra-operator simplifies the management of Apache Cassandra and DataStax Enterprise on Kubernetes, providing a robust and scalable solution for distributed data storage and management in cloud environments.

### Overview: This repository contains the open-source code for the paper "Retrieval Head Mechanistically Explains Long-Context Factuality," authored by [the authors of the paper](https://arxiv.org/abs/2404.15574). The code is based on the implementation of the "Needle In a Haystack" project, which is available at [this repository](https://github.com/gkamradt/LLMTest_NeedleInAHaystack/tree/main). ### Key Features: - **Code Implementation**: The code is implemented based on the "Needle In a Haystack" project, providing a comprehensive solution for detecting retrieval heads in transformer models. - **Model Families**: The code is currently implemented for several model families, including LLaMA, Yi, Qwen, and Mistrial. - **Detection Algorithm**: The algorithm calculates the retrieval score of attention heads in a transformer model, utilizing FlashAttention for decoding. - **Performance**: It supports a maximum length of 50,000 tokens and can be run on a single 80GB GPU, making it suitable for a wide range of applications. - **Usage**: The code provides a simple command-line interface for detecting retrieval heads, allowing users to set the start and end positions of the model and save the results in a JSON file. - **Visualization**: The code includes visualization tools for better understanding the results, with examples provided for both top-K and random masking scenarios. ### Tech Stack: The primary language used is Python, and the code

### Overview: This repository contains a comprehensive list of publically available audio data for Automatic Speech Recognition (ASR) and other speech-related activities. The data is sourced from various projects and platforms, including LibriSpeech, TED-LIUM, VoxForge, Common Voice, Tatoeba, and more. The data is categorized into free and paid options, with sizes ranging from 100 hours for LibriSpeech to 500 hours for VoxForge. ### Key Features: - **LibriSpeech - Train**: 100 hours of training data from LibriSpeech. - **LibriSpeech - Test**: 360 hours of test data from LibriSpeech. - **TED-LIUM Release 2**: 118 hours of audio from TED-LIUM. - **TED-LIUM Release 3**: 452 hours of audio from TED-LIUM. - **VoxForge English**: 130 hours of audio from VoxForge. - **Common Voice v1**: 500 hours of audio from Common Voice. - **Tatoeba Audio Eng**: 200 hours of audio from Tatoeba. - **Valentini Noisy Speech Database**: 10,87 hours of audio from Valentini. - **VOiCES Download Script**: 15 hours of audio from VOiCES. - **NPTEL2020 Indian English Speech Dataset**: 15,700 hours of audio from NPTEL2020. - **Open St

Overview: Texel Studio is an AI pixel art tool that paints like a real artist, using shapes, noise, and per-pixel control instead of diffusion models. It offers exact, consistent, and game-ready results. Key features: - **Realistic Pixel Art**: Paints like a human artist, using shapes, noise, and per-pixel control. - **Sprite Types**: Supports both tileable blocks and transparent background items. - **Concept Art Reference**: AI generates a reference image before pixel painting. - **Live Streaming**: Watch the sprite build in real-time via SSE. - **Chat Continuation**: Send follow-up edits to the same agent session. - **Manual Pixel Editing**: Click to paint, right-click to erase. - **Palette Management**: Create, edit, and reuse color palettes. - **Autotile Generation**: Generate all 16 edge variants for tilemap use. - **Export Options**: Native PNG, upscaled preview, or full tileset folder. - **History**: Browse and reload past generations. - **S3 Storage**: Optional S3-compatible object storage for shared file access across workers. - **Observability**: Optional LangSmith and/or PostHog LLM analytics tracing. Tech stack: Python Why it matters: Texel Studio offers a unique approach to pixel art generation, providing precise control and consistent results, making it an essential tool for artists and developers looking to create high-quality, game-ready pixel art. Repository: EYamanS/texel-studio GitHub stars: 239 GitHub description: AI pixel

### Overview: Discover AI is an interactive desktop app designed to teach modern AI concepts through hands-on labs. It features 51 small, engaging labs covering fundamental AI topics such as machine learning fundamentals, transformers, RAG, agents, fine-tuning, and more. ### Key Features: - **51 Interactive Labs**: Navigate through 51 labs covering various AI concepts. - **Hands-On Play**: Move sliders, click words, and train models to observe changes in real-time. - **Lesson Explanations**: Short, simple, and technical explanations next to each lab. - **Project Management**: Track progress and complete labs for a capstone project. ### Tech Stack: TypeScript ### Why It Matters: Discover AI simplifies complex AI concepts by providing practical, visual demonstrations. It makes advanced ideas accessible through interactive play, encouraging hands-on learning and experimentation.

### Overview: This repository is dedicated to the survey of LLM-based human-agent collaboration and interaction systems, which has been accepted to ACL 2026. It aims to compile an extensive collection of awesome resources relevant to this rapidly evolving field, fostering further research and innovation. ### Key Features: - 📖 Latest Research Papers - 📚 Applications, Datasets & Benchmarks - 📖 Web Navigation & Computer Use - 📖 Software Engineering, Coding - 📖 Embodied AI, Robotics - 📖 Conversation System - 📖 Data Science, Scientific Discovery - 📖 Gaming - 📖 Finance - 📖 Healthcare, Medicine - 📖 General-Purpose Assistants, Cross-Domain - 📖 Retail, Telecom - 📖 Travel - 📖 Writing ### Tech Stack: Python, LaTeX, Markdown, Git, and various research papers and resources. ### Why it Matters: This project aims to provide a comprehensive overview of the latest research in LLM-based human-agent collaboration and interaction systems, highlighting the challenges and solutions in this rapidly evolving field. The contributions are welcome to expand the repository and improve the survey paper.