Projects

Overview: This Elixir client library provides a REST interface to interact with Qdrant, a vector search engine. It supports creating collections, upserting points, performing queries, and managing snapshots. Key features: - REST-based API for Qdrant - Support for creating collections and upserting points - Ability to perform queries and retrieve results - Provision of collection aliases and optimization progress - Support for cluster and service operations Tech stack: The client is built using Elixir, leveraging HTTP/REST for communication. Why it matters: This library simplifies integration with Qdrant, making it easier for developers to leverage vector search capabilities within their Elixir applications.

### Overview: KVBoost is a practical toolkit designed to accelerate local LLM inference by leveraging content-addressed KV cache reuse. It optimizes performance without requiring model porting, making it compatible with existing stacks. ### Key features: - **Chunk-level KV reuse**: Keeps previous requests' prefix-aligned chunks in a cache, allowing subsequent requests to skip redundant work. - **CacheBlend seam repair**: Selectively recomputes at chunk boundaries to maintain output quality, even at high reuse rates. - **KV quantization**: Optional 8-bit or 4-bit cache for 2-4x memory savings with minimal accuracy loss. - **FlashAttention-2 CUDA kernel**: Custom tiled-softmax kernel for Volta → Hopper GPUs. - **AWQ layer streaming**: Runs 32B-class models on 12GB GPUs by streaming INT4 layer weights from pinned host RAM. - **Speculative decoding**: Small AWQ draft proposes K tokens; streamed target verifies in one forward, enhancing performance. ### Tech stack: Python ### Why it matters: KVBoost optimizes LLM inference by reducing redundant work, making it significantly faster while maintaining accuracy. It works seamlessly with existing stacks, ensuring minimal changes to models, tokenizers, or calling code.

### Overview: AutoDub - Automatic Video Translation & Dubbing AutoDub is a powerful tool for converting videos into different languages using AI-powered text-to-speech. It supports 100+ languages and offers three TTS engines for seamless translation and preservation of the original video's audio. ### Key Features: - **Speech Recognition**: Transcribes audio using OpenAI Whisper. - **Translation**: Translates videos to 100+ languages via Google Translate. - **Three TTS Engines**: - **Edge TTS**: High-quality Microsoft voices (recommended). - **Silero**: Fast Russian TTS (offline after first download). - **XTTS**: Voice cloning from 6-10 second samples. - **Video Preservation**: Keeps original video, mixes original audio (20%) with dubbed audio (150%). ### Tech Stack: Python AutoDub is built using Python, leveraging libraries such as OpenAI Whisper, Google Translate, and TTS engines. ### Why It Matters: AutoDub simplifies the process of translating and dubbing videos, making it accessible to a wide audience. Its ability to preserve the original video's audio and provide high-quality translations makes it a valuable tool for creators and language enthusiasts alike. ### GitHub Stars: 31 AutoDub has gained significant traction on GitHub, indicating its growing popularity and usefulness in the community. ### Primary Language: Python AutoDub is primarily developed and maintained in Python, ensuring a robust and versatile solution for video translation and dubbing.

### Overview: Welcome to the Multimodal AI Essentials code repository! This repository provides a comprehensive guide on how to merge text, image, and audio for smarter models. The project includes several Jupyter notebooks that cover various aspects of multimodal AI, such as multimodal semantic search, visual question answering, and diffusion models. The notebooks are designed to help you understand and implement these techniques effectively. ### Key Features: - **Multimodal Semantic Search**: Using the SigLip model to build an image search system. - **Visual Q/A**: Emulating the process of Llama 3.2-Vision-Instruct, including the construction and training of a VQA model. - **Diffusion Models**: Exploring techniques like Stable Diffusion and Dreambooth for generating images and fine-tuning models. - **Text to Speech**: Fine-tuning text-to-speech models to convert text into speech. ### Tech Stack: - Jupyter Notebook ### Why it Matters: This repository is essential for anyone looking to enhance their understanding of multimodal AI and its practical applications. It provides a solid foundation in integrating text, image, and audio for smarter models, making it a valuable resource for developers and researchers in the field of artificial intelligence.

Overview: A simple web application designed to transcribe audio files into text, utilizing Python for the core functionality. Key features: - Audio file upload - Automatic transcription - Real-time progress display - User-friendly interface Tech stack: Python Why it matters: This project provides a straightforward solution for audio transcription, making it accessible for developers and beginners alike.

### Overview: KodiSharp is a C# library that allows developers to create rich Kodi addons using the .NET framework. It provides a seamless integration with Kodi's Python APIs, eliminating the need for temporary files and ensuring robust variable persistence. ### Key Features: - **Python Interoperability**: Uses `ctypes` to interact with .NET modules. - **Events Support**: Handles `xbmc.Monitor` and `xbmc.Player` events. - **Python Interfaces**: Supports code evaluation, variable management, value escaping, and Python console logging. - **C# Bindings**: Provides bindings for Kodi modules such as `xbmc`, `xbmcgui`, etc. - **URL Routing**: Enables defining and connecting different plugin functionalities to menu entries. - **Service Addons**: Runs background activities without blocking the Kodi process. - **Background Execution**: Addons can run in the background, persisting across script invocations. ### Tech Stack: - **.NET Runtime**: Hosted by CLRHost (Windows) or MonoHost (Windows/Mac/Linux). ### Why It Matters: - **Eliminates Temporary Files**: No longer need to use temporary files for variable persistence. - **Robust Variable Persistence**: Variables are kept alive until the host process terminates. - **Background Execution**: Run plugins quickly by recycling instances, ensuring background activities can persist. - **Performance**: Uses an in-process JSON RPC for Python interoperability, reducing unnecessary transitions. - **Code Optimization**: Optimized for frequent calls to Kodi or Python APIs to minimize transitions. ### Performance: - **Code Evaluation**:

Overview: THREEPIO is a real-time, AI-powered conversational droid that can be interacted with through voice, showcasing a unique approach to AI interaction by removing it from the screen and placing it into a physical object. Key features: - Real-time audio pipeline - Ambient interaction loop - Barge-in / interruption handling - Echo and feedback suppression - Persona-driven response layer - Embedded deployment on Raspberry Pi - Modular provider architecture - Hardware + software integration Tech stack: Python Why it matters: This project explores the implications of AI interaction when it is removed from screens and placed into a physical object, addressing the challenges of latency, echo, and barge-in in voice-based interactions, while maintaining a consistent conversational style.

Overview: This project focuses on fine-tuning a BERT model for question answering using a limited dataset for illustration purposes. Key features: - Fine-tuning a BERT model for question answering - Utilizing a limited dataset for training - Achieving satisfactory results on a test dataset Tech stack: SimpleTransformers for BERT model fine-tuning Why it matters: This project demonstrates the potential of fine-tuning BERT for question answering tasks, even with a small dataset, highlighting the model's ability to extract contextual information from input text.

### Overview: Sagalee: Automatic Speech Recognition Dataset for Oromo Language The Sagalee dataset is an open-source Automatic Speech Recognition (ASR) dataset specifically designed for the Oromo language. This dataset aims to provide researchers and developers with a comprehensive resource for training and evaluating ASR models on the Oromo language. ### Key Features: - **Language:** Oromo Language - **Dataset Type:** Automatic Speech Recognition Dataset - **License:** CC BY-NC 4.0 International - **GitHub Stars:** 31 ### Tech Stack: - Python ### Why It Matters: The Sagalee dataset offers a valuable resource for researchers and developers working on ASR for the Oromo language. It provides a structured dataset that can be used to train and evaluate ASR models, contributing to advancements in language processing and speech recognition technology. ### Repository: - **Repository:** turinaf/Sagalee - **Primary Language:** Python - **GitHub Stars:** 31 ### Access: - **OpenSLR:** [https://openslr.org/157/](https://openslr.org/157/) - **Huggingface:** [turinaf/Sagalee](https://huggingface.co/datasets/turiabu/Sagalee) ### Training ASR on Sagalee Dataset: - **Clone Repo:** ```bash git clone https://github.com/turinaf/sagalee.git cd sagalee git submodule update --init --no-fetch

Overview: Weaviate CLI is a powerful command-line interface designed for managing and interacting with Weaviate vector databases directly from the terminal. Key features: - Collections: Create, update, delete, and get collection configurations. - Data Management: Import, query, update, and delete data with various search types (vector, keyword, hybrid). - Multi-tenancy: Manage tenants and their states across collections. - Backup & Restore: Create and restore backups with support for S3, GCS, and filesystem. - Sharding: Monitor and manage collection shards. - Flexible Configuration: Configure vector indexes, replication, consistency levels, and more. - Role Management: Assign and revoke roles and permissions to users. Tech stack: Python Why it matters: Weaviate CLI simplifies the management of Weaviate vector databases, making it easier for developers to interact with the database directly from the terminal, enhancing productivity and efficiency in data management tasks. Repository: weaviate/weaviate-cli GitHub stars: 31 GitHub description: CLI tool for Weaviate

### Overview: SMART is an innovative multi-agent framework designed to internalize complex trajectories in knowledge-intensive tasks using Long- and Short-Trajectory Learning. This general paradigm can be expanded to other complex tasks, enhancing the capabilities of arbitrary multi-agent frameworks. ### Key features: - Long-Trajectory Learning - Short-Trajectory Learning - Trajectory Dataset Construction - Training Scripts - Evaluation Datasets ### Tech stack: Python ### Why it matters: SMART provides a unified framework for internalizing trajectories in knowledge-intensive tasks, making it a valuable tool for researchers and developers working on complex AI applications.

### Overview: SEIKO is a reinforcement learning method designed to efficiently fine-tune diffusion models for high-reward regions on the feasible manifold. This approach, presented in the paper "Feedback Efficient Online Fine-Tuning of Diffusion Models," outperforms traditional baselines in image, biological sequence, and molecular fine-tuning tasks. ### Key features: - **SEIKO (Optimi**s**tic fin**E**-tuning of d**I**ffusion with **K**L c**O**nstraint)**: A novel reinforcement learning method for fine-tuning diffusion models. - **Efficient Exploration on Feasible Samples**: Targets high-reward regions on the manifold of feasible samples. - **Theoretical and Empirical Validation**: Validated across three domains: images, biological sequences, and molecules. ### Tech stack: Python ### Why it matters: SEIKO significantly improves the efficiency and effectiveness of fine-tuning diffusion models, particularly in high-reward regions, by leveraging reinforcement learning to discover optimal samples. This method is particularly useful in applications where the goal is to maximize certain properties, such as aesthetic quality or bioactivity, on a constrained manifold.

### Overview: Eagle-RAG is an enterprise RAG platform designed to enhance AI applications by integrating semantic knowledge and visual insights. It leverages Knowhere, PixelRAG, and LlamaIndex to provide intelligent search capabilities, making it easier to find relevant information quickly. ### Key Features: - **Semantic Knowledge Integration**: Seamlessly combines semantic chunks with pixel-native perception to enhance search accuracy. - **Multi-tenant Data Layer**: Built to support multiple industries and teams, ensuring data isolation and efficient retrieval. - **Microkernel + In-Repo Plugins Architecture**: Core provides basic functionalities, while domain plugins extend hooks and encoders. - **Experimental and Under Development**: Features like `plugins/biomed` and `plugins/lakehouse_bi` are currently in development. - **Built-in UI**: A showcase of Knowhere and PixelRAG, suitable for both developers and non-developers. - **Backend + MCP for Downstream Agents**: Supports vertical-specific knowledge bases and multiple data layers. ### Tech Stack: - Python 3.12 - Milvus, LlamaIndex, Knowhere, PixelRAG, and MinerU ### Why It Matters: - **Semantic Depth and Visual Clarity**: Search knowledge by understanding both the meaning and the visual aspects of documents. - **Multi-tenant Support**: Built to handle multiple industries and teams without data mixing. - **Efficient Retrieval**: Integrates multiple collections and ANN systems for comprehensive search results. ### Repository: - Repository: zhiweio/EagleRAG - GitHub Stars: 30

Overview: This is a sample Unity project designed to demonstrate the integration of Speech Recognition and Language Understanding using the new Microsoft Speech Service (Preview) and LUIS from Microsoft Cognitive Services. Key features: - Speech Recognition - Language Understanding with LUIS Tech stack: Unity C# Why it matters: This project serves as a practical example for developers looking to integrate speech recognition and language understanding capabilities into their Unity projects using the latest Microsoft services. Repository: ActiveNick/Unity-SpeechWithLUIS GitHub stars: 30 GitHub description: Sample Unity project used to demonstrate the integration of Speech Recognition and Language Understanding using the new Microsoft Speech Service (Preview) and LUIS from Microsoft Cognitive Services. This is a work in progress.

Overview: mini-DeepInfer 是一个面向 Ascend 910C 的精简教学推理引擎，使用真实 Qwen3.5 权重实现，通过一条请求来展示请求调度、模型执行、采样和结果提交的过程。 Key features: - 实现 DeepInfer 执行设计 - 使用精简的 Python 与 PyTorch - CPU 和 Ascend 910C 两种平台验证机制 - 提供十四课学习路径 - 源码对应映射及双模型逐轮演示 Tech stack: Python, PyTorch, NPU, CANN Why it matters: 通过 mini-DeepInfer，开发者可以快速学习 DeepInfer 的源码实现，理解请求调度、模型执行、采样和结果提交的流程，从而更好地掌握 DeepInfer 的设计和实现。

Overview: This repository aims to create an Automatic Speech Recognition model for the Arabic language using two prominent ASR frameworks: Kaldi and CMU-Sphinx4. The dataset used is the "Arabic Corpus of Isolated Words," consisting of 10,000 utterances of 20 words spoken by 50 native male Arabic speakers, available for free download. Key features: - Arabic ASR model using Kaldi - Arabic ASR model using CMU-Sphinx4 - Dataset: "Arabic Corpus of Isolated Words" with 10,000 utterances of 20 words, 50 speakers, and 20 words per speaker, totaling 10,000 audio files Tech stack: Shell Why it matters: This project provides a practical demonstration of using two leading ASR frameworks for Arabic language, offering insights into the implementation and performance of these tools on real-world data.

Overview: SenTrEv is a Python package designed to evaluate the performance of Sentence Transformers on text retrieval tasks, specifically tailored for dense and sparse retrieval on PDF documents. Key features: - Evaluates Sentence Transformers models for dense and sparse retrieval on PDFs. - Supports dense and sparse text encoders from the `sentence_transformers` and `fastembed` libraries. - Handles various document types including PDF, PPTX, DOCX, HTML, CSV, and XML. - Provides a simple evaluation workflow with customizable settings. Tech stack: Python Why it matters: SenTrEv simplifies the process of evaluating Sentence Transformers models for text retrieval tasks, making it easier to choose the best model for RAG applications with text-based documents, and it offers a straightforward evaluation workflow with customizable parameters. Repository: AstraBert/SenTrEv Primary language: Python GitHub stars: 30 GitHub description: Simple customizable evaluation for text retrieval performance of Sentence Transformers embedders on PDFs

Overview: Use Atlas Cloud's 300+ AI models inside ComfyUI. Drop-in nodes for Sora 2, Veo 3.1, Kling 3, Seedance 2, Nano Banana Pro, GPT Image 2, Flux 2 & more. Key features: - Atlas Cloud nodes for various AI models - Support for Sora 2, Veo 3.1, Kling 3, Seedance 2, Nano Banana Pro, GPT Image 2, Flux 2, and more - Drop-in nodes for easy integration into ComfyUI - Supports 300+ AI models Tech stack: Python Why it matters: This project allows developers to easily integrate cutting-edge AI models into their ComfyUI projects, enhancing their workflows and capabilities without the need for local GPUs or model weights.

### Overview: LongTrainer is a production-ready RAG framework that turns your documents into intelligent, multi-tenant chatbots with 5 lines of code. ### Key Features: - **Multi-tenant bots**: Isolated bots for each tenant. - **Persistent MongoDB memory**: Encrypted and restorable. - **FAISS vector search**: Fast and accurate search. - **Streaming responses**: Yield chunks of responses. - **Custom tool calling**: Plug-and-play support. - **Chat encryption**: Secure chat communications. - **Vision support**: Pass images for vision chat. - **Self-improving from chats**: Feeds Q&A back into the KB. ### Tech Stack: Built on LangChain. ### Why It Matters: LongTrainer simplifies the process of building production-ready RAG systems, making it easy to integrate and manage multi-tenant chatbots with all the necessary features and capabilities.

### Overview: LongParser is a privacy-first document intelligence engine designed for production RAG pipelines. It parses PDFs, DOCX, PPTX, XLSX, and CSV files into AI-ready chunks, including validated, semantic, and cross-referenced information. The system includes a Human-in-the-Loop (HITL) review process and a three-layer memory architecture for enhanced recall and context. ### Key Features: - **Multi-format extraction**: Supports PDF, DOCX, PPTX, XLSX, and CSV files. - **Hybrid chunking**: Token-aware, heading-hierarchy-aware, and table-aware chunking. - **Semantic chunking**: Embedding-based boundaries using `all-MiniLM-L6-v2`. - **Cross-referencing**: Deterministic linking of explicit and implicit charts/figures. - **Quality scoring**: Zero-ML heuristic scoring with dictionary validation. - **PII redaction**: Hybrid Regex + NER (spaCy) redaction with secure HITL preservation. - **Summary chunks**: Async ARQ worker generating hierarchical LLM section summaries. - **HITL review**: Human-in-the-Loop block and chunk editing before embedding. - **LangGraph HITL**: `approve / edit / reject` workflow with LangGraph `interrupt()` and MongoDB checkpointer. - **3-layer memory**: Short-term turns + rolling summary + long-term facts. - **Multi-provider LLM**: OpenAI, Gemini, Groq, OpenRouter. - **Multi-backend vectors**: Chroma, FAISS, Qdr