Projects

Overview: `encoder-run` is a Kubernetes operator designed to automatically generate vector embeddings for source code repositories. It ensures that embeddings stay current with the latest changes in code, enhancing search, similarity checks, RAG, and code analysis tasks. Key features: - Continuous Synchronization: Monitors source code repositories and triggers the encoding process for updated code, ensuring embeddings remain current. - Model Deployment: Deploys SentenceTransformer/Encoding models to process embeddings internally. - Scalable and Efficient: Optimized for performance and scalability, handling multiple repositories and large codebases with ease. - Customizable Encoding Models: Supports a variety of encoding models, allowing users to choose the best fit for their specific requirements. Tech stack: TypeScript Why it matters: This project ensures that vector embeddings stay up-to-date with the latest changes in source code repositories, enhancing various applications that rely on code similarity and analysis. Repository: encoder-run/operator Primary language: TypeScript GitHub stars: 52 GitHub description: Kubernetes operator for producing vector embeddings based on source code repositories.

### Overview: Toy-scale experiments exploring the "Neo-Unify" hypothesis, a single transformer backbone that can unify image understanding and generation without degrading either task. ### Key features: - **Neo-Unify (Phase C)**: A unified model with shared attention and task-specific expert FFNs for both understanding and generation. - **Synthetic 16x16 RGB Images**: 18,000 images across 6 classes for Neo-Unify variants and 6,000 for Phase A/B baselines. - **MLX on Apple Silicon**: All models use NHWC format per MLX convention. - **Incremental Approach**: Build through 5 phases, each adding capability or testing a hypothesis. ### Tech stack: Python 3.9 ### Why it matters: Explore whether a single transformer can unify image understanding and generation without degrading either task, supporting the "Neo-Unify" hypothesis.

### Overview: This project is a PyTorch implementation of a transfer learning approach for automatic lyric transcription (ALT) using wav2vec 2.0, showcasing significant improvements over previous methods. ### Key features: - Transfer learning of wav2vec 2.0 for ALT - Adaptation of wav2vec 2.0 to the singing domain - Exploration of different transfer starting points - Hybrid CTC/attention model for enhanced performance - Significant performance over previous approaches on various ALT benchmark datasets ### Tech stack: Python ### Why it matters: This project addresses the performance gap between automatic speech recognition (ASR) and automatic lyric transcription (ALT) by leveraging the similarities between speech and singing through transfer learning. The results demonstrate that even with a small amount of training data, the method can achieve competitive performance, making it a valuable contribution to the field of automatic lyric transcription.

### Overview: A developer-focused project for generating and processing text-to-speech models, featuring enhanced inference capabilities and support for various model architectures and features. ### Key features: - **Flow Matching TTS** - Rectified Flow Diffusion Transformer (RF-DiT) over continuous DACVAE latents. - **Multi-version Model Support** - One codebase for v1/v2/v3/v2 & v3-VoiceDesign/v4-Small/v4.1-Small checkpoints. - **Quantized Checkpoint Loading** - torchao-quantized v4-Small-Quantized checkpoints. - **Voice Cloning** - Zero-shot voice cloning from reference audio. - **Voice Design (Caption Conditioning)** - Describe voice style/emotion in natural language. - **Duration Predictor** - Automatic output length on supported models. - **Speaker Inversion (v4 family)** - Trainable speaker tokens. - **Pretrained Text Encoder (v4 family)** - Loads pretrained backbone + projector. - **MeanFlow Distillation Training (v4.1)** - `train_mode: meanflow_distill` distills an RF-trained checkpoint. - **Emoji Style Control** - Emotion and sound-effect control via emojis. - **Emotion Style Presets** - One-click presets with fine-grained CFG control. - **Speaker Library** - Register reference audio as named speaker profiles. - **Multi-Candidate Generation** - Up to 8 candidates per run. - **Multiline Split Generation** - Per-line individual output or concatenation with silence gaps. - **Water

Overview: This project serves as an inference stack for Nvidia DGX Spark, providing a way to run large language models (LLMs) on a desktop setup. The stack is primarily based on vLLM models and is designed for the less affluent. Key features: - Models start on the first request and hold the GPU until idle for 20 minutes. - A small utility model can remain loaded for light tasks like session titles. - The project is a hobby project for a single DGX Spark setup and has not been tested with two-node setups. Tech stack: Node.js 20 or later, Docker, NGC account for base images. Why it matters: This project provides a practical solution for running large language models on a desktop, making it accessible to those who may not have the resources for dedicated AI infrastructure.

Overview: ASRecognition is an easy-to-use library for Automatic Speech Recognition that utilizes pre-built components from Hugging Face. It simplifies adding speech recognition features to applications. Key features: - Transcribes audio files to text - Supports multiple languages including Arabic, German, Greek, English, Spanish, Persian, Finnish, French, Hungarian, Italian, Japanese, Dutch, Polish, Portuguese, Russian, and Chinese Tech stack: Python Why it matters: This library provides a straightforward interface for integrating speech recognition capabilities into applications, making it accessible for developers without extensive coding knowledge. Repository: jonatasgrosman/asrecognition GitHub stars: 52

Overview: Cascadia is a model distribution system that allows running any language model on Intel hardware, eliminating the need for cloud or NVIDIA GPUs. Key features: - OpenAI-compatible API - Pipeline parallelism - Built-in sharder - Seven engines - Single static binary per node - Zero-config peer discovery - `cascadia doctor` for diagnostics Tech stack: Rust Why it matters: Cascadia enables efficient model distribution and inference on Intel hardware, making it a cost-effective and scalable solution for running large language models without the need for expensive cloud services or GPUs. Repository: labscommunity/cascadia GitHub stars: 52 GitHub description: Run any model on Intel silicon

### Overview: This project provides an end-to-end solution for speech recognition, text translation, and text-to-speech functionalities on iOS using Amazon Translate and Amazon Polly as AWS Machine Learning managed services. It showcases the power of machine learning in mobile applications with a simple and straightforward implementation. ### Key features: - Speech Recognition - Text Translation - Text-to-Speech ### Tech stack: Swift ### Why it matters: This solution demonstrates the integration of AWS services into iOS applications, making it a valuable tool for developers looking to incorporate machine learning into their projects. The simplicity of the implementation and the ease of use make it a practical choice for beginners and experienced developers alike.

### Overview: This project is a benchmark for speech-to-text engines, comparing Velma Transcribe, Deepgram, AssemblyAI, Chirp 3, MAI-Transcribe 2, Moonshine, and whisper.cpp on LibriSpeech, VoxPopuli, and Common Voice. The benchmark is reproducible and uses the same audio and scoring criteria. ### Key features: - **LibriSpeech, VoxPopuli, Common Voice**: 20 English files per dataset, 163 to 482 reference words per dataset. - **Metrics**: Word Error Rate (WER), Latency, and USD per audio hour. - **Engines**: Velma Transcribe, Deepgram, AssemblyAI, Chirp 3, MAI-Transcribe 2, Moonshine, whisper.cpp. - **Hosted APIs**: Velma, Deepgram, AssemblyAI, Chirp 3, MAI-Transcribe 2. - **Local Engines**: Moonshine, whisper.cpp. ### Tech stack: Python ### Why it matters: This benchmark provides a standardized way to compare the performance of speech-to-text engines, allowing developers to easily evaluate their models against a set of well-defined criteria.

Overview: The Moorcheh Python SDK is a powerful tool for interacting with the Moorcheh Semantic Search API, offering ultra-fast, highly accurate vector similarity search and analysis capabilities. Key features: - Ingest raw text, files, or vectors with a unified API - One-shot RAG (Relevance-Augmented Generation) for quick answers - Zero-ops scale, serverless architecture that scales up and down automatically - Ideal for production-grade RAG, agentic memory, and semantic analytics Tech stack: Python Why it matters: The Moorcheh SDK simplifies the process of integrating Moorcheh's advanced semantic search capabilities into Python applications, making it an essential tool for developers looking to enhance their AI applications with high-performance, accurate search and analysis. GitHub stars: 52 Repository: moorcheh-ai/moorcheh-python-sdk

Overview: OneLLM is a unified interface for interacting with various LLMs from hundreds of providers, offering intelligent semantic caching, fallback mechanisms, and enhanced reliability. Key features: - Supports hundreds of models from dozens of providers - Provides a consistent API for OpenAI-compatible patterns - Includes built-in fallback mechanisms - Supports 22 providers with over 300 models Tech stack: Python Why it matters: Simplifies integration of LLMs into applications, making it easier to use multiple models and providers with a familiar interface. Repository: muxi-ai/onellm GitHub stars: 52 GitHub description: Unified interface for interacting with various LLMs hundreds of models, caching, fallback mechanisms, and enhanced reliability.

Overview: ASRT Speech Recognition Java SDK is a comprehensive Java-based SDK designed for ASRT voice recognition systems, offering a robust solution for developers looking to integrate speech recognition capabilities into their applications. Key features: - Java-based SDK - Supports Java 8+ - Compatible with Eclipse development environment - Deployed on Tomcat or Nginx for web integration - Demonstrates a demo client for quick development and integration - Compatible with any web project, embedding backend functionalities Tech stack: Java Why it matters: This SDK simplifies the integration of speech recognition technology into Java-based applications, making it an essential tool for developers working with ASRT systems. Repository: nl8590687/ASRT_SDK_Java GitHub stars: 52 GitHub description: ASRT Speech Recognition Java SDK. 用于ASRT语音识别系统的Java SDK

### Overview: PEFT-SingLoRA is an efficient alternative to traditional LoRA that uses a single low-rank matrix instead of two, reducing parameters while maintaining performance. This package provides a PEFT-compatible implementation of SingLoRA for the Hugging Face ecosystem, making it easy to integrate with existing PEFT workflows. ### Key Features: - **50% fewer parameters** than standard LoRA. - **Fully compatible with PEFT** ecosystem. - **Mathematically equivalent** to standard LoRA. - **Easy integration** with existing PEFT workflows. ### Tech Stack: Python ### Why it Matters: PEFT-SingLoRA simplifies the process of low-rank adaptation by reducing the number of parameters needed, making it a valuable tool for improving the efficiency and performance of transformer models without sacrificing accuracy.

Overview: This project contains a collection of PocketBase plugins designed to enhance functionality and add new features. Key features: - Full Text Search - Vector Search/Store - WebRTC - ENV Config Tech stack: Go Why it matters: These plugins provide developers with additional tools and features to optimize their PocketBase applications, enhancing user experience and functionality.

### Overview: This repository contains code for the [O'Reilly Live Online Training for Deep Learning for Modern AI](https://learning.oreilly.com/live-events/deep-learning-for-modern-ai/0642572005084). The training provides foundational knowledge for machine learning and deep learning using PyTorch, focusing on practical concepts and real-world case studies. The repository includes several Jupyter Notebook examples that cover topics such as object recognition, text classification, fine-tuning models, and optimization strategies. ### Key features: - Introduction to Deep Learning - Fine-tuning BERT and Llama 3 for various applications - Production Optimization and Quantization Techniques - Multimodal Models with CLIP and SHAP-E + Diffusion - Visual Q/A Case Study ### Tech stack: - Jupyter Notebook ### Why it matters: This training provides essential theoretical knowledge and practical skills for building and optimizing machine learning models, particularly those used in modern AI applications such as image recognition, text classification, and generative models. The use of large pre-trained models and advanced optimization techniques ensures that models are both efficient and performant. ### Repository: - **Primary language**: Jupyter Notebook - **GitHub stars**: 52 - **GitHub description**: Code for Deep Learning for Modern AI - **README excerpt**: 

### Overview: Vechord is a Python library that enables PostgreSQL to be used as a search engine, providing vector, multivec, and keyword search functionalities with advanced features like contextual retrieval, embedding generation, and more. ### Key Features: - **Vector Search with RaBitQ**: Uses VectorChord for efficient vector search. - **Multivec Search with WARP**: Utilizes VectorChord for multivec search. - **Keyword Search with BM25 Score**: Provides keyword search with BM25 scoring. - **Auto-Generated Web Service**: Simplifies development with a web service generator. - **Data Consistency**: Ensures data consistency through PostgreSQL transactions. - **Decorator for Data Injection**: Easily injects data into the PostgreSQL tables. - **Common Tools**: Offers tools like `Augmenter`, `Chunker`, `Embedding`, `Evaluator`, `Extractor`, `EntityRecognizer`, `Reranker`, and `GroundTruth`. - **Examples**: Provides examples for various use cases, including simple, flexible, web-based, essay, contextual, hybrid, graph, and dynamic pipelines. ### Tech Stack: Python ### Why It Matters: Vechord simplifies the integration of PostgreSQL as a search engine, making it easier to build powerful search applications with Pythonic practices. ### Repository: tensorchord/vechord ### GitHub Stars: 52 ### Installation: Simply install Vechord using `pip install vechord`. ### Related Docker Images: Available in DockerHub and GitHub Packages. ### Examples: - `simple.py`: Basic example for

Overview: TinyVecDB is a lightweight, embedded vector database designed for simplicity and speed, focusing on performance and minimal resource usage. Key features: - Embedded: No server setup or maintenance required - Simple API: Quick start with just a few lines of code - Cross-platform: Works seamlessly on Windows, macOS, and Linux - Lightweight: Minimal memory footprint and resource usage - Offline-capable: Ideal for desktop and mobile applications Tech stack: C Why it matters: TinyVecDB offers a fast and efficient solution for vector storage and similarity search, making it suitable for applications requiring high performance and minimal overhead. Repository: tylerpuig/tinyvec GitHub stars: 52 GitHub description: TinyVecDB is an ultra fast embedded vector database.

### Overview: AI Dictation is an open-source speech-to-text app designed for macOS, Windows, iPhone, iPad, and Android. It offers offline recognition, optional cleanup, and personal vocabulary, making it useful for various applications and workflows. ### Key features: - **Offline Recognition**: Supports offline transcription for non-voice input. - **Optional Cleanup**: Allows for post-transcription formatting and cleanup. - **Personal Vocabulary**: Enables users to add their own words for faster recognition. - **Writing Rules**: Allows users to define rules for specific words or phrases. - **Spoken Shortcuts**: Provides spoken alternatives for frequently used words or phrases. ### Tech stack: Swift ### Why it matters: AI Dictation is a versatile tool for anyone who frequently types or writes text, whether it's for work, personal projects, or daily communication. Its ability to recognize speech and provide accurate transcriptions makes it a valuable addition to any developer's toolkit, especially for those who need to quickly capture their thoughts and ideas. ### Repository: writingmate/aidictation ### GitHub stars: 51 ### GitHub description: Open-source speech-to-text and dictation app for macOS, Windows, iPhone, iPad, and Android, with offline recognition and optional cleanup.

Overview: This repository is an official implementation of the ICML 2024 paper "Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts." The paper discusses a tool called P4D that automatically identifies problematic prompts for diffusion models to test the reliability of safety mechanisms. Key features: - P4D is a debugging and red-teaming tool for text-to-image diffusion models. - It finds problematic prompts to test the reliability of safety mechanisms. - The tool uncovers new vulnerabilities in diffusion models with safety mechanisms. Tech stack: The repository uses Python for the implementation. Why it matters: P4D is significant because it provides a practical way to test the reliability of safety mechanisms in text-to-image diffusion models, which can help ensure the security and ethical use of such models.

### Overview: This repository serves a 4-bit Qwen3.8-27B model on a single DGX Spark (GB10 / SM121), achieving speeds of up to 75 tok/s single-stream and 246 tok/s aggregate at 8-way concurrency. The repository includes a side-by-side comparison of two 4-bit checkpoints and the FP8 baseline, measured on a single harness. ### Key Features: - **Speed**: 75 tok/s single-stream (`k=14`), or 246 tok/s aggregate at 8-way concurrency. - **Concurrency**: Single-stream decode is memory-bandwidth-bound, with a 7–9% throughput cost at high concurrency. - **Quantization**: Measured against FP8 on the same harness, DSpark `k=7`: - c1: 46.91 - c4: 134.21 - c8: 208.71 - c16: 256.08 ### Tech Stack: - **Python** used for the repository. ### Why It Matters: - **Quantization Advantage**: NVFP4 (Unsloth) + DSpark `k=14` achieves 72.6–75.0 tok/s, offering a 27% improvement over FP8 on the same harness. - **Concurrency Trade-offs**: The 4-bit model's throughput is limited by the number of weight reads, with higher concurrency