Projects

### Overview: This repository is a curated, research-grade catalog of public AI models that support the Kazakh language. It includes models for various domains such as Text & NLP, Speech & Audio, Vision & OCR, and Watchlist resources that are either announced or not yet independently verifiable. ### Key Features: - **Text & NLP**: - **[Kazakh Trilingual GLiNER](https://huggingface.co/ardakshalkar/kazakh-trilingual-gliner)**: A trilingual (Kazakh, Russian, English) zero-shot Named Entity Recognition model based on GLiNER, open-sourced. - **[Darmm Chat Kazakh 8B](https://huggingface.co/Darmm/darmm-chat-kazakh-8b)**: A conversational assistant for Kazakh, trained on 105k instruction pairs (65k Kazakh) using Qwen3-8B. - **[Darmm Embed Kazakh v2](https://huggingface.co/Darmm/darmm-embed-kazakh-v2)**: A second-generation dense retrieval and text embedding model for Kazakh, fine-tuned from BAAI/bge-m3. - **Speech & Audio**: - **[Kazakh Speech Dataset](https://github.com/Allessyer/awesome-kaz-datasets)**: A curated, research-grade catalog of public Kazakh-capable speech datasets. - **Vision & OCR**: - **[K

Overview: Scowld is an open-source iOS app that provides a live AI companion with a customizable avatar and advanced features such as speech-to-speech, hands-free wake mode, and live voice providers. Key features: - Animated companion with customizable avatar and live voice provider integration - Live speech-to-speech functionality - Interrupt anytime feature - Hands-free wake mode - Home tips for custom key and plan management - BYOK live voice providers (Google Gemini Live, OpenAI Realtime, xAI Grok Voice) - Scowld Premium (optional) for no API key and unlimited access - Optional vision feature - Captions and past chat transcripts - Ends by itself after 90 seconds of silence or leaving the app - Character settings for avatar, name, and prompt Tech stack: Swift Why it matters: Scowld offers a unique and advanced voice companion experience, combining live AI interaction with customizable features and live voice provider integration, making it a valuable tool for users looking for a hands-free and personalized communication assistant. Repository: apoorvdarshan/scowld GitHub stars: 34 GitHub description: 🦫 Open-source iOS AI companion with a live VRM avatar that sees you — bring-your-own-key chat, vision, speech-to-text & text-to-speech (OpenAI, Anthropic, Gemini, ElevenLabs, Groq & more). SwiftUI.

### Overview: End-to-end speech recognition on the AISHELL dataset using Pytorch. ### Key features: - Attention-based sequence-to-sequence model - Bidirectional GRU encoder with BatchNorm - GRU decoder with Luong-based attention - 80-dimensional filter banks as acoustic features - SpecAugment applied to improve generalization - Achieves ~10% CER on test set after 100 epochs ### Tech stack: Pytorch ### Why it matters: This project showcases advanced speech recognition techniques, demonstrating the effectiveness of attention-based models and data augmentation methods in improving speech recognition performance.

### Overview: Gemini Omni is a cutting-edge AI model from Google capable of processing and generating text, code, images, audio, and video seamlessly. This project aims to curate and showcase the best resources, platforms, and creative examples related to this powerful tool. ### Key Features: - **Official Resources**: Comprehensive guidelines and tutorials from Google DeepMind, including official product pages, prompt guides, and model cards. - **Interactive Platforms**: Google Flow, a creative workspace powered by Gemini Omni, enabling interactive collaboration and video editing. - **Capabilities and Showcases**: Native video editing, object insertion and substitution, background replacement, and material color alteration. ### Tech Stack: - **GitHub**: Repository for this project. ### Why it Matters: Gemini Omni is a revolutionary tool that democratizes AI capabilities across various domains, making it accessible for developers and creatives alike. Its ability to seamlessly integrate text, code, images, and video is particularly noteworthy, offering a versatile platform for creative expression and problem-solving.

Overview: AgilePruner is an empirical study that investigates the strengths and weaknesses of attention-based and diversity-based pruning methods in large vision-language models. The project aims to provide insights into how these approaches handle visual token sequences and feature diversity, with the goal of improving the performance of adaptive visual token pruning. Key features: - Conducts empirical analysis using effective rank and attention score entropy as measures. - Reveals that diversity-aware hybrid pruning methods can lead to increased hallucination frequency compared to attention-based pruning. - Highlights that attention-based approaches are more effective on simple images, while diversity-based methods are better suited for complex images. Tech stack: Python Why it matters: AgilePruner offers valuable insights into the performance and limitations of different pruning strategies, potentially leading to more effective visual token pruning techniques in large vision-language models.

### Overview: AX Engine is a Mac-first, Apple Silicon inference runtime optimized for Qwen 3.8 27B AXQ. It provides real speed and multi-model support with prefix reuse and MTP acceleration, built for Apple Silicon and measured against MLXLM and LLaMA.cpp. ### Key Features: - **Qwen 3.8 27B AXQ Pinned MXFP4 MTP Pack**: Selects the pinned MXFP4 MTP pack for Qwen 3.8 27B AXQ. - **MXFP8 Pack Admission Pending**: MXFP8 pack admission is pending, with pack qualification and qualification center support. - **MTP Execution Explicit Opt-In Until Gates Pass**: Execution remains explicit opt-in until gates pass. - **Standalone Affine 4/6/8-bit and Non-MTP Variants Retired**: Standalone affine variants and non-MTP variants are retired. - **Standalone Default Load Mode**: The default load mode is `add`, allowing multiple models to be resident. - **MXFP4 and MXFP8 Packed Models**: MXFP4 and MXFP8 packed models are available for qualification. - **MXFP4 MTP Pack Pinned**: MXFP4 MTP pack is pinned for Qwen 3.8 27B AXQ. - **Flash Next 125B-A6B Model**: Supports the Flash Next 125B-A6B model with MXFP4 MTP. - **Pinned MXFP4 MTP

### Overview: This project is an agentic-coding inference engine optimized for NVIDIA V100-SXM2 GPUs, designed to enhance the efficiency and performance of Claude Code and related applications. ### Key features: - **GLM-5.3-Flash**: Fast tool-call turns with a primary model for agentic coding. - **Qwen3.8-Flash-Next**: High-speed tool-call turns on four cards for text and image processing. - **DeepSeek-V4.1-Flash**: The most capable model for reasoning and world knowledge, running locally on V100-SXM2. - **MiniMax-H3**: Fast text-to-video and keyframe generation on four GPUs. ### Tech stack: Python ### Why it matters: This engine is optimized for speed and efficiency, making it ideal for applications requiring quick and accurate responses, especially in agentic coding scenarios where speed is crucial.

### Overview: Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. This project explores how different RAG components interact and how their performance can be evaluated independently. ### Key Features: - **Swap One Stage, Hold the Rest Fixed**: Measure the impact of changing a single stage in a RAG pipeline. - **Compose and Compare**: Combine multiple RAG stages and compare their outputs to determine their relative contributions. - **Corpus Sizes**: Tested with corpora ranging from 5K to 200K documents. ### Tech Stack: Python ### Why It Matters: This project provides a practical way to evaluate and compare RAG components, demonstrating how different stages of a pipeline can be isolated and their effects measured. This is crucial for understanding and optimizing the compositionality of RAG systems, which can significantly impact their performance and effectiveness.

Overview: The project aims to create and merge the best Stable Diffusion models, including both SD1.5 and SDXL variants. Key features: - Fluently XL Final - Fluently XL v4 - Fluently XL v3 - Fluently XL v3 Lightning - Fluently XL v3-inpainting - Fluently XL v2 - Fluently XL v1 - Fluently v4 - Fluently v4-inpainting - Fluently v4-LCM - Fluently v3.5 - Fluently v3 - Fluently v3-inpainting - Fluently Epic - Fluently Anime - Fluently v2.5 - Fluently v2 - Fluently v1 Tech stack: Python Why it matters: The project contributes to the development of the Stable Diffusion model family, providing high-quality images and enhancing the community's involvement in the project.

Overview: A Telegram/Nostr bot for AI image generation with Bitcoin Lightning payments. Key features: - Supports Stable Diffusion and DALLE 2 - Payments and refunds via Bitcoin Lightning Network - Multiple AI models (DALLE2 and Stable Diffusion) - Lightning payments and refunds with LNbits - Logging functionality - Multiple chat protocols (Nostr and Telegram) Tech stack: Python Why it matters: This bot is a valuable tool for those looking to generate high-quality images using AI with secure and transparent payment methods, making it an essential resource for developers and content creators. Repository: f321x/lightningpicturebot Primary language: Python GitHub stars: 34 GitHub description: Telegram/Nostr bot for AI image generation with Bitcoin Lightning payments README excerpt:

Overview: This repository contains professional skills for building, auditing, evaluating, and optimizing RAG (Retrieval-Augmented Generation) systems with Claude Code. It includes features such as RAG Audit, RAG Eval, Chunking Advisor, and RAG Scaffold, which help developers improve their RAG pipeline quality and efficiency. Key features: - RAG Audit: Analyzes existing RAG code for anti-patterns and issues. - RAG Eval: Evaluates RAG quality with metrics and benchmarking. - Chunking Advisor: Provides recommendations for optimal chunking strategies. - RAG Scaffold: Generates production-ready RAG boilerplate. Tech stack: The skills are implemented using Python and LangChain. Why it matters: These skills are crucial for developers looking to build and optimize RAG pipelines with Claude Code, ensuring their systems are efficient, reliable, and production-ready.

### Overview: This repository contains scripts to create a merged version of multiple open-source German speech datasets, addressing the lack of large speech corpora for automatic speech recognition tasks. The merged corpus is stored in `data/full`, and it can be used with the `audiomate` library for further processing. ### Key Features: - Scripts to download, merge, and convert multiple German speech corpora into a single, unified dataset. - Corpus statistics for different parts, including speakers and utterances. - Corpus format compatible with the `audiomate` library for easy reading and processing. - Corpus sources listed, including Common-Voice, Mailabs, SWC, TUDA, and VoxForge. ### Tech Stack: Jupyter Notebook ### Why it Matters: This repository fills a gap in the availability of German speech corpora for automatic speech recognition tasks, making it essential for researchers and developers working on German language speech recognition systems.

Overview: 一个基于 NVIDIA Parakeet-tdt-0.6b 模型的本地语音转录服务，提供与 OpenAI API 兼容的接口和简洁的 Web 用户界面，适用于音视频文件的快速转录和生成高精度的 SRT 字幕。 Key features: - 实时语音转录 - 简洁易用的 Web 用户界面 - 支持多种语言，包括英语、日语、越南语等 - 与 pyVideoTrans 等现有工作流无缝集成 Tech stack: Python Why it matters: 该服务提供了一个高效、准确且易于使用的解决方案，适用于需要高质量字幕的视频制作和教育场景，同时节省了对互联网 API 调用的依赖，降低了成本和风险。

Overview: See & Sniff: Learning Visuo-Olfactory Representations is an official PyTorch implementation of a paper presented at ECCV 2026. The project focuses on training models to learn visuo-olfactory representations, which are crucial for various applications in computer vision and robotics. Key features: - Training of models for visuo-olfactory representation learning - Implementation of a global-CLIP model - Evaluation of models using checkpoints and evaluation scripts - Data preparation instructions for datasets Tech stack: The project uses Python and PyTorch for development and training. It also utilizes Hugging Face datasets for dataset preparation. Why it matters: This project is significant for researchers and practitioners interested in developing models that can effectively extract and utilize information from both visual and olfactory inputs, which has applications in various fields such as robotics, computer vision, and environmental sensing. Repository: kaistmm/SeeandSniff Primary language: Python GitHub stars: 34 GitHub description: [ECCV 2026 Oral] Official Pytorch implementation for "See & Sniff: Learning Visuo-Olfactory Representations"

### Overview: Q-ASR is an innovative approach for zero-shot quantization of ASR models, enabling efficient speech recognition without the need for training or validation data. This method utilizes integer-only computation to achieve significant performance improvements while maintaining high accuracy. ### Key Features: - **Zero-Shot Quantization**: Q-ASR does not require any training or validation data, focusing solely on integer-only computation. - **Efficient Performance**: Achieves state-of-the-art results on Librispeech benchmark, demonstrating superior performance over traditional quantization methods. - **Versatile Application**: Suitable for various ASR models, including QuartzNet15x5Base-En, demonstrating broad applicability across different ASR systems. ### Tech Stack: - **Programming Languages**: Python - **Dependencies**: Cython, NeMo ### Why It Matters: Q-ASR represents a significant advancement in ASR quantization, offering a practical solution for reducing computational complexity and resource requirements without compromising on performance. Its ability to achieve high accuracy with minimal data and computation makes it particularly valuable for real-world applications where efficiency and resource constraints are critical.

Overview: This repository contains the code for a paper presenting a novel method called IAIS, which aims to improve cross-modal retrieval by aligning textual and visual relations. Key features include a Relation Consistency Hypothesis and a novel metric called ISDa, and a new regularized training method called IAIS. Key features: - Relation Consistency Hypothesis - Novel ISDa metric - New IAIS regularized training method Tech stack: Python Why it matters: The IAIS algorithm achieves a new state-of-the-art (SOTA) performance on the Winoground dataset, improving upon previous methods by 10% on VinVL and 52% on UNITER, demonstrating significant advancements in cross-modal retrieval.

### Overview: ArionTalk is an open-source voice AI agent that allows users to interact with websites by voice. It automatically understands the content of a page, highlights relevant sections, and provides natural turn-taking behavior. The agent runs on-device with Chrome's built-in AI or in the cloud with the Gemini Live API. ### Key Features: - **Page Understanding:** Reads the text, headings, and images on the page before answering. - **Interactive Highlights:** Scrolls to and highlights the section or image being discussed. - **Natural Turn-Taking:** Stops to listen when interrupted. - **Two Engines, One API:** Runs on-device with Chrome's Prompt API, or in the cloud with the Gemini Live API. ### Tech Stack: - TypeScript ### Why It Matters: ArionTalk simplifies voice interaction with websites, making it accessible and intuitive for users to ask questions about content without needing a backend or server.

Overview: This project demonstrates how to use the Google Cloud Speech API to transcribe audio/video files. Key features: - Transcribes audio files from mp4 format - Converts audio to FLAC encoding - Breaks the audio file into 10-second clips - Transcribes each clip and prints the result to the console Tech stack: PHP Why it matters: This project provides a practical guide for developers looking to integrate Google Cloud Speech API into their applications for audio transcription, ensuring efficient and accurate speech-to-text conversion.

### Overview: Helix AI is a powerful AI workspace that allows you to chat, create agents, code, and manage knowledge bases on your own machines. It offers a seamless experience with local models, ensuring no data leaves your control. ### Key Features: - **Local Models:** Helix AI runs models on your machine or through a local server, using LM Studio's engine. It avoids cloud model selection. - **Chat and Agents:** You can have conversations and create agents directly within the app. - **Coding:** Helix AI supports coding, allowing you to write and run code snippets. - **Knowledge Bases (RAG):** It includes a knowledge base that can be fine-tuned with local models. - **Fine-Tuning:** Helix AI allows fine-tuning of models with local models, ensuring the best performance for your environment. ### Tech Stack: - TypeScript ### Why It Matters: Helix AI provides a secure, local AI workspace where all data remains under your control, ensuring privacy and security. It eliminates the need for cloud model selection, making it ideal for teams and organizations that value data sovereignty.

### Overview: StreamLingo is a real-time AI-translated captions extension for YouTube livestreams, translating between English and Chinese. It supports local Ollama and automatically detects the source language, ensuring accurate translations without requiring captions from the streamer. ### Key features: - **Real-time AI-translated captions**: StreamLingo generates Chinese or English translations directly from livestream audio, displayed inside the player. - **Local speech recognition**: Uses local speech recognition and the Ollama model you already run. - **Automatic language detection**: Supports English → Chinese, Chinese → English, and auto-detection of either Chinese or English output. - **No streamer captions needed**: The extension captures audio from the browser tab and translates it directly. - **Windows/NVIDIA preview**: Supports Windows and NVIDIA, with a prebuilt Windows preview that starts with a double-click without manual commands. ### Tech stack: Python 3.11–3.13 ### Why it matters: StreamLingo simplifies the process of live captioning, ensuring accurate translations without the need for external captions or manual setup. Ideal for anyone looking to improve their live streaming experience by providing real-time translations in their preferred language.