Projects

### Overview: This project is a Question Answering application utilizing Large Language Models (LLMs) and Amazon OpenSearch Service. It leverages the Retrieval Augmented Generation (RAG) approach to retrieve and bundle relevant information from the enterprise knowledge base or content, enhancing the accuracy of the LLM responses. ### Key Features: - Question Answering (QA) application - Integration with Amazon OpenSearch Service - Retrieval Augmented Generation (RAG) approach - Deployed using Amazon SageMaker and Amazon OpenSearch Service - Access credentials stored in AWS Secrets Manager - Streamlit application for easy deployment ### Tech Stack: Python ### Why it Matters: This project demonstrates the effectiveness of integrating large language models with Amazon OpenSearch Service, specifically using the RAG approach to improve the accuracy of responses. It showcases how enterprises can leverage cutting-edge AI technologies to enhance their knowledge management and question answering capabilities.

### Overview: This repository is a curated list of AI resources, including courses, tools, and GitHub repositories, designed to help beginners and experienced practitioners alike understand and utilize Generative AI, Large Language Models, and Agentic AI. ### Key Features: - **Projects**: A collection of projects and repositories focused on Generative AI, LLMs, and AI agents. - **Free Courses**: Access to free courses on AI development, integration, and fluency. - **Awesome Lists**: A curated list of AI tools and frameworks. - **Books**: A selection of AI-related books. - **AI/ML Tools & Frameworks**: A comprehensive list of AI and ML tools and frameworks, including AI Gateway, AI Workload Manager, Copilot Development, Dataset Engineering, Evaluation, Fine Tuning, Function Calling, Graph RAG, Guardrails, Local Model Inference, LLM Agent Framework, Model Serving, Observability, Pre Training, Prompt Engineering, RAG Framework, Security, Structured Extraction, Structured Generation, Vector DB. ### Tech Stack: GitHub ### Why it Matters: This cheat sheet serves as a comprehensive guide for anyone interested in Generative AI, Large Language Models, and Agentic AI, providing essential resources and tools to enhance their understanding and practical application of these technologies. Repository: bhanuchaddha/ai-cheat-sheet GitHub Stars: 31

### Overview: Effortless Push-to-Talk Transcription, Anywhere. ### Key features: - **User-Friendly Interface**: Directly configure input device, model, compute type, and language. - **CPU-friendly**: Fastest on any CPU with **[cutting-edge ASR models](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard)**. - **LLM correction**: Optional repair of broken transcriptions using any OpenAI-compatible API. ### Tech stack: Python ### Why it matters: Transcribes spoken text into text-based chat, making it easy to use in any environment.

Overview: This project is a Python script designed to automatically generate subtitle and translated subtitle files for any video or audio file using the faster_whisper module and an unofficial online Google Translate API. Key features: - Automatically generates subtitle and translated subtitle files - Supports multiple languages for subtitles and translations Tech stack: Python Why it matters: This script streamlines the process of creating subtitles for videos and audio files, making it easier for users to watch or listen to content without needing to manually add captions. GitHub stars: 31

Overview: SmartRAG is a production-ready Multi-Modal RAG (Reinforcement Learning with Attention) engine designed for querying databases, documents, images, and audio in natural language. It supports multi-AI and multi-database coordination, making it versatile for various applications. Key features: - Unified query across databases, documents, images, and audio - Supports multi-AI providers (e.g., OpenAI, Anthropic) - In-memory storage for efficient querying - Production-ready with vector storage and multi-database coordination - NuGet package available for .NET 6 Tech stack: C# Why it matters: SmartRAG simplifies complex querying tasks by integrating multiple sources of information, making it an invaluable tool for developers and businesses looking to enhance their data retrieval capabilities in a multi-modal environment. Repository: byerlikaya/SmartRAG GitHub stars: 31 GitHub description: Multi-Modal RAG for .NET — query databases, documents, images and audio in natural language. Production-ready with multi-AI support, vector storage, and multi-database coordination.

Overview: This repository contains scripts designed to streamline the data generation, scoring, and preparation for the CHiME-8 DASR task. It provides a convenient interface for downloading and preparing the core CHiME-8 data, including CHiME-6, DiPCo, and NOTSOFAR1, as well as ESPNet/Kaldi-style manifests. Key features: - Generates and downloads CHiME-8 data - Prepares CHiME-8 data lhotse manifests - Prepares CHiME-8 data ESPNet/Kaldi-style manifests - Scripts for official scoring Tech stack: Python Why it matters: This package simplifies the process of obtaining and preparing the necessary data for the CHiME-8 DASR task, making it easier for developers to integrate and evaluate their models. Repository: chimechallenge/chime-utils GitHub stars: 31

Overview: HyperVectorDB is a local vector database built in C# that supports various distance/similarity measures, including Cosine Similarity, Jaccard Dissimilarity, Euclidean, Manhattan, Chebyshev, and Canberra distances. Key features: - Supports multiple distance/similarity measures - Query/Response Caching - Cache invalidation - Query Functions: Cosine Similarity, Jaccard Dissimilarity, Euclidean, Manhattan, Chebyshev, Canberra Distance - Automatic Parallelization - Data Compression Tech stack: C# Why it matters: HyperVectorDB is designed for applications that require high-performance vector queries, making it suitable for use in recommendation systems, content-based filtering, and other scenarios where vector similarity is important. Repository: deatos/HyperVectorDB GitHub stars: 31

### Overview: KOSR is a Korean speech recognition project based on transformer, providing end-to-end Korean speech recognition models and a dataset for training. This project aims to improve the accuracy of speech recognition tasks in the Korean language. ### Key Features: - **Transformer Joint CTC (Testing)**: A model that combines transformer with connection-to-connection training. - **Transformer Transducer**: A model that uses a transformer architecture for speech recognition. - **Transformer**: A model that processes speech data using a transformer architecture. ### Tech Stack: Python ### Why it Matters: KOSR is designed to enhance the accuracy of speech recognition in the Korean language, making it useful for applications such as voice-activated assistants, voice-over-IP (VoIP) systems, and other voice-based communication tools. The project includes models that can be trained and tested, providing a comprehensive solution for speech recognition tasks in Korean. ### Repository: - **Repository**: dobby-seo/kosr - **GitHub Stars**: 31 - **GitHub Description**: Korean speech recognition based on transformer (트랜스포머 기반 한국어 음성 인식) ### Preparation: - **Dataset**: Download the KsponSpeech dataset from AI-Hub and prepare the directory structure as shown in the README. - **Environment**: Requires Python 3.6 or higher, PyTorch 1.7.0 or higher, and torchaudio 0.7.0 or higher. - **Training**: Use the provided scripts and configurations to train the models.

### Overview: 小红书内容收集器（XHS Collector）是一款Chrome扩展，旨在通过被动采集、多模态AI分析以及本地存储功能，帮助用户高效管理和分析小红书笔记。它结合了本地IndexedDB存储、多模态AI分析、MCP服务器检索和CLI工具箱，提供了一站式的笔记收集、整理和分析解决方案。 ### Key features: 1. **被动采集与分析**：通过浏览器插件收集小红书笔记，并进行多模态AI分析，包括爆款拆解、仿写、标签建议和视觉诊断。 2. **本地存储与导出**：本地IndexedDB存储笔记，支持多种格式导出，方便后续分析和使用。 3. **MCP服务器集成**：通过MCP服务器检索本地语料，提供实时只读工具，如xhs_status、xhs_search、xhs_read、xhs_comments和xhs_hot。 4. **多格式支持**：兼容多种笔记格式，如JSON、JSONL、Markdown和训练数据。 5. **实时分析工具**：CLI工具箱提供实时搜索、读详情、评论和热门内容功能，支持OpenRouter、Anthropic、Google AI等模型。 6. **防风控与隐私保护**：内置防风控机制，确保用户数据安全，同时限制访问频率和敏感信息。 ### Tech stack: - **Node.js**: 用于构建和运行MCP服务器。 - **IndexedDB**: 本地存储解决方案，用于去重和持久化笔记数据。

Overview: This project is a demonstration of using large language models (LLMs) to interact with local data, showcasing the concept of retrieval-augmented generation (RAG). It involves embedding the data, using similarity search to find relevant chunks, and combining them with the question to generate an answer. Key features: - Embedding model - Similarity search - Vector database - Persistent storage Tech stack: Python Why it matters: This project serves as an educational tool to understand how LLMs can be used with local data, demonstrating the practical application of RAG in a real-world scenario. Repository: fau-masters-collected-works-cgarbin/gpt-all-local Primary language: Python GitHub stars: 31 GitHub description: A "chat with your data" example: using a large language models (LLM) to interact with our own (local) data. Everything is local: the embedding model, the LLM, the vector database. This is an example of retrieval-augmented generation (RAG): we find relevant sections from our documents and pass it to the LLM as part of the prompt.

### Overview: Ollama Herd is an open-source smart AI router that turns your existing machines (Mac, Linux, and Windows PCs) into a single endpoint, allowing you to route models like LLMs, embeddings, image generation, and speech-to-text across them. It supports Claude Code, Codex, Open WebUI, and any OpenAI- or Ollama-compatible clients, making it a versatile solution for developers. ### Key Features: - **Smart Scoring**: Routes to the best machine based on thermal state, memory fit, queue depth, latency, affinity, availability, and context fit. - **Zero-Config Discovery**: Auto-discovery via mDNS, no IPs, no config files, no manual setup. - **Multimodal Routing**: Routes LLMs, vision, embeddings, image generation, and speech-to-text. - **Live Dashboard**: Fleet overview, trends, model insights, per-app analytics, benchmarks, health, recommendations, settings. - **Capacity Learning**: 168-slot weekly behavioral model per device, learns when your machines are idle. ### Tech Stack: Python ### Why It Matters: - **Every machine you own, one endpoint**: Macs, Linux servers, and Windows PCs pooled into one fleet. - **Agents stop queuing behind one Ollama**: Each request goes to the best machine for that model right now. - **Claude Code and Codex on your own hardware**: Model IDs auto-route to the best model you have loaded, no model map needed. - **No config files,

Overview: This project is a speech recognition implementation using Connectionist Temporal Classification (CTC) with Keras and TensorFlow backend. It includes training logs and saved models for various neural network architectures. Key features: - Speech recognition with RNN and CTC - Implementation using Python 2.7, TensorFlow 1.6.1, and Keras 2.1.5 - Supports installation on macOS, Ubuntu, and Windows - Includes training logs and saved models for different architectures Tech stack: Python, TensorFlow, Keras Why it matters: This project demonstrates the effectiveness of recurrent neural networks and connectionist temporal classification for speech recognition, providing a framework for further research and development in the field.

Overview: This project involves creating a voice-activated assistant, inspired by J.A.R.V.I.S. from the Marvel Cinematic Universe, using Python and the ChatGPT API on a Raspberry Pi. Key features: - Voice recognition using PyAudio - Text-to-speech conversion using pyttsx3 - Integration with the ChatGPT API for speech synthesis Tech stack: Python Why it matters: This project demonstrates the practical application of AI and voice technology, providing a unique and engaging experience for users. Repository: jasmineplows/jarvis GitHub stars: 31 GitHub description: Using ChatGPT API and a Raspberry Pi to create my own J.A.R.V.I.S. from the Marvel Cinematic Universe Blog post: https://medium.com/@jasmineplows/how-to-build-your-own-version-of-iron-mans-voice-activated-assistant-j-a-r-v-i-s-using-chatgpt-4b738bbe8147

Overview: Wave Subs is a subtitle generator that translates and generates SRT/ASS subtitles for videos, using local AI and no internet. It supports multiple languages and is free and open-source. Key features: - Automatically generates subtitles from videos - Supports various subtitle formats and languages - Runs entirely on your local machine - Recognizes dialogue and aligns it with speech timing - Translates subtitles into your preferred language Tech stack: TypeScript Why it matters: Wave Subs simplifies the process of finding subtitles for videos, making it accessible and free without requiring internet access or the need for an account.

### Overview: VoiceAssistant is a locally hosted voice assistant with a plugin extension feature, utilizing the VOSK speech recognition engine. It allows for easily configurable commands via JSON configs and supports multiple languages, low CPU load, and various plugins for tasks such as playing music, checking weather, and more. ### Key Features: - **Locally Hosted Voice Recognition and Speech Generation** - **Plugin Extensibility Model** - **Easily Configurable Commands via JSON Configs for Each Plugin** - **Multi-Language Support (Change Settings for Speech Generation and Model of Speech Recognition, then Rewrite Plugin Commands to Your Language)** - **Almost No CPU Load on Low Power Intel Core i5 3rd Generation. Should Even Work on Intel Atom** - **Available Plugins:** Hello, Timer, Open Web-Page, Run Program, Currency Rates, Weather Check, Google Calendar Tasks Check, Application Control using Key Code Injection, Play Music, Message Broadcast, Voice Connection, MQTT Adapter, Telegram Audio Message Receiver ### Tech Stack: - C# ### Why It Matters: - Locally hosted speech recognition and generation, making it accessible and reliable. - Plugin extensibility allows for easy integration of various functionalities. - Supports multiple languages and configurations. - Low CPU load, ensuring efficient performance even on low-powered devices. ### GitHub Stars: 31 ### Available Plugins: - Hello - Timer - Open Web-Page - Run Program - Currency Rates - Weather Check - Google Calendar Tasks Check - Application Control using Key Code Injection - Play Music

Overview: Nala is a fastapi voice-assistant framework designed to quickly prototype large-language-model-powered voice assistants in just 5 minutes. It integrates with state-of-the-art transcription and text-to-speech engines, making it easy to integrate with LLMs like Dolly-v2-3b. Key features: - Extensible architecture for easy integration with new response models and TTS voice skins. - Direct integration with Dolly-v2-3b LLM model. - Cross-platform support for web applications, mobile apps, and IoT devices. - Audio-to-Audio API for submitting and getting audio responses. - Simple user interface for rating responses. - Privacy and security features including admin downloads and user authentication. Tech stack: Python, FastAPI, Flask, Web Speech API, Dolly-v2-3b, Whisper API, SpeechT5 model. Why it matters: Nala provides a streamlined and efficient way to prototype voice assistants, making it easier for developers to integrate with advanced AI models and create engaging user experiences.

### Overview: llambda.lisp is an independent Common Lisp inference engine for running quantized Large Language Models directly from `.gguf` files. ### Key features: - **Native GGUF Parsing:** Directly ingests and parses `Q4_K_M` and `Q6_K` quantized tensors from disk. - **AVX2 / FMA Acceleration:** The core GEMV (Matrix-Vector Multiplication) bottleneck is pulverized using SBCL's `sb-simd`. - **Multi-Threaded Execution:** The outer loop of the GEMV processing is fully parallelized via `lparallel`, saturating modern multi-core memory buses. - **Zero-Drift KV Cache:** Safe, shared-KV reuse and perfectly aligned RoPE scaling. - **Advanced Sampling:** Built-in Top-K, Top-P (Nucleus), and repetition penalties executing in-place with zero heap allocation in the hot path. - **Experimental Ryzen AI NPU backend:** Fixed-shape ONNX projections can replace selected CPU matrix-vector operations through an optional CFFI bridge. - **Experimental DirectML GPU backend:** Selected projections can run on DirectML-capable AMD, Intel, or NVIDIA GPUs. ### Tech stack: Common Lisp ### Why it matters: Because the industry has succumbed to the dogma that C++ is the only path to bare-metal AI inference. `llambda.lisp` exists to prove that properly architected, aggressively typed, and hardware-aware Common Lisp can achieve C-level throughput without sacrificing the interactive, REPL-driven elegance of Lisp. ### Repository:

Overview: This repository is a comprehensive database of China policy shocks, providing crucial data for economic, financial, and policy evaluation research. Key features: - Contains detailed records of policy changes and their impacts. - Includes historical data spanning multiple years. - Facilitates easy access and analysis for researchers. Tech stack: Python Why it matters: This resource is essential for researchers looking to understand the effects of policy changes in China, offering valuable insights into economic and financial dynamics.

Overview: LibDriver LD3320 is a full-featured driver library designed for general-purpose MCU and Linux, providing voice recognition, MP3 playback, and additional features. Key features: - Voice Recognition - MP3 Playback - Dynamic Key Editing - Integration with Induction Cooker, Microwave Oven, Smart Home Appliances, and More Tech stack: C Why it matters: LibDriver LD3320 simplifies the integration of speech recognition functionality into existing products, making it easier to add voice recognition, voice control, and human-computer dialogue capabilities without the need for additional hardware components. GitHub stars: 31 Repository: libdriver/ld3320

### Overview: This project is a Semantic QA system that queries markdown files using vector embedding, Pinecone vector database, and GPT (langchain). It converts markdown files into vectors and uploads them to Pinecone, allowing for efficient and fast retrieval of relevant information. ### Key Features: - Queries any markdown file using vector embedding and Pinecone vector database. - Works with Notion and Obsidian, converting markdown files to `.md` format automatically. - Utilizes Feynman technique for practice. - A weaker version of privateGPT and llama_index, suitable for practice or small-scale projects. ### Tech Stack: - Pinecone vector database - OpenAI embedding model - Langchain for text processing and vector embedding ### Why It Matters: This project is a practical application of vector embedding and natural language processing, demonstrating how to efficiently query and retrieve information from markdown files. It serves as a useful tool for developers looking to improve their documentation or query management systems.