Projects

### Overview: Hybrid RAG is a Python package designed to streamline the Retrieval-Augmented Generation (RAG) pipeline, making it easy to track, evaluate, and optimize your system for real-world applications. ### Key Features: - **Ingestion Pipeline**: End-to-end support for inserting vector data into Milvus VectorDB. - **Experiment Tracking & Tracing with MLflow**: Log experiments, parameters, and traces for every LLM and retrieval step, ensuring efficient latency & cost tracking. - **RAG Evaluation with Ragas**: Measure performance using faithfulness, answer relevance, and context precision, with future support for MLflow evaluation. - **Cost Monitoring**: Keep track of API usage by setting LLM pricing inside API parameters to optimize expenses. - **Hybrid Search Capability**: Semantic (dense) & keyword (sparse) retrieval, query expansion, Milvus-optimized retrieval, self-query retrieval, reranking, and auto-metadata filtering. - **Nemo Guardrails (v0.1.1)**: Uses vector similarity for question classification, reducing middleware time, preventing prompt injection attacks, and enforcing policy restrictions. - **Smart Summarization & Q&A Handling**: Supports direct QA over documents, metadata filtering, and map-reduce summarization for extracting insights across document chunks. - **Follow-up Question Generation**: Auto-generate follow-up questions to improve engagement with users. - **Custom PyFunc Hybrid-RAG MLflow Model**: Register, deploy, and serve the best model directly as an MLflow API for production-grade scenarios. - **Optimized Modules with Async

### Overview: TrooperAI is a low-latency, local voice assistant built for the Raspberry Pi 5, featuring real-time speech recognition, LLM-based dialog, and high-quality TTS. It integrates a Game 5 Pi retro arcade case, LED feedback, and gesture control, and utilizes the Gemma3 and Qwen2.5 models for a responsive and realistic interaction. ### Key Features: - **Fully Integrated into Raspberry Pi 5**: Runs on stock 8GB RAM and 32GB MicroSD card. - **No Remote Dependencies**: Operates independently, avoiding cloud providers. - **WebSocket Architecture**: Supports full-duplex mic/speaker communication. - **Sentence Streaming STT**: Uses lightweight Vosk model for speech recognition. - **Sentence-by-Sentence TTS**: Uses Piper for text-to-speech with customizable voices. - **LLM Inference**: Locally processed using Ollama models. - **Configurable Mic-Mute Mode**: Supports setup with separate mic and speaker. - **LED Feedback**: Provides visual feedback through large LED buttons. - **Gesture Detection**: Alternative to mic detection for gesture-based control. - **Performance**: CPU maxed out at 100% during testing, with active cooler fan and additional case fan. ### Tech Stack: Python ### Why it Matters: The TrooperAI project demonstrates the feasibility of building a low-latency, local voice assistant using Python and Raspberry Pi, showcasing the potential for integrating AI into physical devices without reliance on remote services or cloud infrastructure.

Overview: This repository contains the official implementation of CMA, a method for enhanced out-of-distribution detection through cross-modal alignment of multi-modal representations. The project is based on a paper presented at CVPR 2025, highlighting the importance of fine-tuning multi-modal representations for better out-of-distribution detection. Key features: - Cross-Modal Alignment (CMA) - Fine-tuning of multi-modal representations - Evaluation on ImageNet-1k OoD benchmark datasets Tech stack: Python Why it matters: The project demonstrates the effectiveness of fine-tuning multi-modal representations for improved out-of-distribution detection, surpassing existing methods in terms of state-of-the-art performance and leading ID accuracy.

### Overview: MultiBanana is a challenging benchmark designed for evaluating multi-reference text-to-image generation models. The dataset includes 3,769 tasks across 36 task types, covering varying reference numbers (up to 8), domain and scale mismatches, multilingual text rendering, and rare concepts. The benchmark aims to serve as a foundation for future research in this area. ### Key features: - **Multi-reference settings**: Varying the number of references (up to 8), domain and scale mismatches among references (e.g., photo vs. anime), and multilingual text rendering. - **Rare concepts**: Including rare concepts like red banana. - **Evaluation framework**: Uses Qwen3-VL-8B-Instruct as a fixed, open-weight judge model (GPT and Gemini are also supported). ### Tech stack: - Python ### Why it matters: MultiBanana provides a comprehensive benchmark for evaluating multi-reference text-to-image generation models, ensuring that models can effectively incorporate information from multiple reference images. This benchmark is crucial for advancing research in this field and improving the accuracy and diversity of generated images.

### Overview: The Unified Multilingual Dataset of Emotional Human Utterances is a comprehensive dataset containing 87,364 audio files of emotional human utterances. These files have been trimmed of leading silences and encoded as PCM signed 16-bit little-endian .wav files, sampled at 16 kHz with a sample width of 16 bits and a bit rate of 128 kbps. The dataset includes labels for speaker gender, emotion category, and valence, as well as unique speaker IDs and duration. ### Key features: - **Audio Files**: 87,364 audio files of emotional human utterances. - **Labels**: Speaker gender, emotion category, and valence. - **Duration**: 258,332.5 seconds (71.76 hours) of audio. - **Samples**: 87,146 samples. - **Speakers**: 1,307 unique speakers. - **Languages**: 9 languages (English, non-English, and non-speech). ### Tech stack: The dataset is built using Jupyter Notebook. ### Why it matters: This unified dataset provides a valuable resource for researchers and developers interested in studying emotional human communication across multiple languages and cultures. It offers a rich source of data for various applications, including natural language processing, emotion recognition, and cross-linguistic research.

### Overview: Turkish Speech Recognition using Facebook's Wav2vec 2.0 models. ### Key features: - Fine-tuned Wav2vec 2.0 models for Turkish speech recognition. - Models achieve high accuracy on Common Voice test splits (8.83% WER on 6.1 TR, 8.62% WER on 7 TR, 10.61% WER on 8 TR). - Preprocessed datasets for training, including text normalization and handling of specific corpora. - Available on the Hugging Face model hub for fine-tuning and evaluation. ### Tech stack: Python ### Why it matters: This project provides a practical solution for Turkish speech recognition using advanced pre-trained models, making it useful for various applications requiring speech-to-text translation and analysis.

Overview: This is a library designed to integrate the Web Speech API with Angular, providing functionalities for speech synthesis and recognition. Key features: - Speech synthesis module for text-to-speech functionality - Speech recognition service for continuous speech recognition - Confidence filtering and alternative retrieval operators Tech stack: TypeScript Why it matters: This library simplifies the integration of speech capabilities into Angular applications, making it easier to implement speech-based features without needing to dive into the complexities of Web Speech API directly. Repository: ng-web-apis/speech Primary language: TypeScript GitHub stars: 32 GitHub description: A library for using Web Speech API with Angular

Overview: This repository contains a fully functional Automatic Speech Recognition system for Unity using DeepSpeech, which can be integrated into any Unity project. Key features: - Automatic Speech Recognition - DeepSpeech integration - Cross-platform support - C# implementation Tech stack: C# Why it matters: This package provides a practical solution for developers looking to add speech recognition functionality to their Unity projects, making it easier to build applications that can understand spoken commands or feedback.

### Overview: ASIST is a real-time assistant designed for Mac and Windows, capable of answering aloud, showing your schedule and emails as cards, and handing over longer work tasks to an agent. It runs on a model chosen from Anthropic, OpenAI, Google, or Cerebras. ### Key Features: - **Real-time Answering:** Starts answering with a short phrase as soon as you finish, and stops reading aloud when you speak over it. - **Card-Based Interaction:** Weather, events, and emails appear as cards beside the conversation. - **Mini Apps:** Agent jobs, tasks, notes, mail, memory, and the calendar open from the Dock or conversation. - **Handing Over Work:** Research and file edits go to the codex or claude CLI, and you can watch the progress and open the results. - **Memory and Diary:** Every day it sorts the day's conversations into its memory of you, and writes a diary of its own. - **Support for Multiple Languages:** The interface and conversation history support eleven languages. ### Tech Stack: TypeScript ### Why It Matters: ASIST is a tool designed to make your daily tasks easier and more efficient. It listens to your voice, remembers your conversations, and handles tasks like research and file edits with an agent, all while keeping your data safe on your computer.

Overview: piqc is a free, read-only GPU waste scanner for Kubernetes inference clusters, designed to help identify idle GPUs, oversized GPU tiers, and wasted capacity in vLLM deployments. Key features: - Finds idle GPUs - Identifies oversized GPU tiers - Detects wasted capacity Tech stack: Python Why it matters: piqc provides a fast and efficient way to identify and address GPU waste in Kubernetes clusters, helping organizations optimize their spending and improve resource utilization. Repository: paralleliq/piqc Primary language: Python GitHub stars: 32 GitHub description: Free, read-only GPU waste scanner for Kubernetes inference clusters, by Paralleliq. Finds idle GPUs, oversized GPU tiers and wasted capacity in vLLM deployments.

Overview: The Terraform Provider for Pinecone allows Terraform to manage Pinecone resources such as indexes, collections, projects, API keys, service accounts, users, invites, and role bindings. Key features: - Supports both regular and admin operations for Pinecone resources. - Provides secure authentication using environment variables or provider configuration. - Enables Terraform to automatically manage Pinecone resources without manual configuration. Tech stack: Go Why it matters: This provider simplifies the management of Pinecone resources in Terraform, making it easier to automate infrastructure and ensure consistency across environments.

### Overview: This project is a collection of examples and development tools built on top of the Hugging Face candle library, powered by Rust. It includes real-time voice processing and inference capabilities for AI models, aiming to create a VTuber-like AI-driven character. The examples are designed to be used for both development and demonstration purposes. ### Key Features: - **Silero VAD (from files)**: Real-time Voice Activity Detection (VAD) using Silero models. - **Silero VAD (realtime without duration threshold)**: Real-time VAD implementation with minimum duration threshold. - **Silero VAD (realtime like ChatGPT voice chat)**: Real-time VAD with ChatGPT-like voice chat integration. - **Silero VAD + Whisper (realtime like ChatGPT voice chat)**: Combining Silero VAD with Whisper for real-time voice chat. - **Whisper Realtime (from microphone)**: Real-time voice processing using Whisper model. ### Tech Stack: Rust, Rust WebAssembly (RWS), DuckDB WASM, PGLite, and candle library from Hugging Face. ### Why It Matters: This project aims to bridge the gap between AI and multimedia technologies, enabling real-time voice processing and inference capabilities for AI-driven VTubers. It provides a comprehensive set of tools and examples to help developers create and improve AI-driven virtual characters. ### Repository: The project is part of the Project AIRI, a group of talented individuals focused on developing AI-driven projects. It is also associated with the Project AIRI,

### Overview: Splish is a fork of Inco's Splash, designed specifically for Apple M5 GPUs. It was built on top of Splash 1.1.0 and offers optimized kernels tailored for a 40-core M5 Max, focusing on Qwen3.8-27B and its fine-tunes. Splish aims to provide better performance and accuracy compared to the original Splash. ### Key Features: - **Optimized Metal Kernels:** Splish includes tuned kernel choices for the M5 Max, improving performance on Qwen3.8-27B and Qwen3.6-35B-A3B. - **Custom Table Tuning:** A new table has been tuned for the 20-core M5 Pro. - **Non-Latin Tokenization Fixes:** The copy rule has been updated to handle non-Latin characters more effectively. - **Benchmark Tools:** The tools used for benchmarking are available upstream if needed. ### Tech Stack: C++ ### Why It Matters: Splish is a one-man project focused on optimizing the performance of Qwen3.8-27B on Apple M5 GPUs. It offers a more efficient and accurate solution compared to the original Splash, making it a valuable resource for developers working with Qwen-based models.

### Overview: This project provides a fine-tuning tool for the MXNet framework, specifically designed for macOS and Linux. The tool supports training on CPU and GPU, including support for the Inception, ResNet, SqueezeNet, and MobileNet models. It eliminates the need for Docker installation and simplifies the process by automatically handling data preparation, model download, and training. ### Key features: - Automatic creation of RecordIO data from images. - Support for multiple GPUs (e.g., eGPU, CUDA). - Fine-tuning of pre-trained models for various datasets. - Easy configuration through a `config.yml` file. - Automatic generation of training accuracy graphs and confusion matrices. - Support for uploading results to Slack for collaboration and monitoring. ### Tech stack: Python ### Why it matters: This tool simplifies the fine-tuning process for researchers and developers working with MXNet, making it accessible without the need for Docker and providing a straightforward configuration system. The ability to fine-tune pre-trained models on multiple GPUs ensures efficient and effective training, particularly useful for complex tasks like image classification.

Overview: A machine learning-powered, voice-based virtual assistant designed for Raspberry Pi, supporting features like conversation, weather, opening websites, geolocation, date/time, timers, and more. Key features: - Geolocation - Weather (temperature, humidity, etc.) based on current location - Date/time features (day, month, year, etc.) - Opening websites (Google, GitHub, Wikipedia, etc.) in the browser - Creating, pausing, and deleting timers - Repeating a certain phrase a specified number of times - Exiting on command Tech stack: Python Why it matters: This assistant provides a powerful and versatile solution for automating tasks and enhancing user experience on a Raspberry Pi, making it ideal for home automation, smart home applications, and more. Repository: reybahl/assistant Primary language: Python GitHub stars: 32 GitHub description: A machine learning powered and speech-based virtual assistant specifically designed for Raspberry Pi. Supports several features like conversation, weather, opening websites, geolocation, date/time, and creating timers.

### Overview: This project focuses on the classification of Cyrillic Mongolian text using modern TensorFlow 2. The work includes fine-tuning on TugsTugi's Mongolian BERT model and other NLP experiments. The repository contains various classifiers using simple neural networks and fine-tuning on TPU and GPU. ### Key Features: - Fine-tuning of TugsTugi's Mongolian BERT model on TPU. - Fine-tuning of Mongolian BERT on GPU. - Classifiers using simple neural networks. - Pretrained Word2Vec initialization from Facebook's fasttext. - 1D Convolution and multiple 1D convnets. - LSTM and LSTM with Attention. - Classification using Mongolian BERT and Tensorflow 2.0. - Classification using Mongolian BERT large and HuggingFace. - Sequence loss in Keras and TensorFlow 2. - Variational Auto Encoder for Mongolian text. ### Tech Stack: TensorFlow 2 ### Why it Matters: This project showcases the application of modern deep learning techniques in text classification, including fine-tuning on pre-trained models and various neural network architectures. The inclusion of fine-tuning on TPU and GPU demonstrates the potential for efficient computation and improved performance. The use of various classifiers and fine-tuning methods highlights the versatility of the approach and the importance of adapting models to specific use cases.

### Overview: This project leverages the idle NVENC silicon on modern Nvidia GPUs to compress intermediate model state, such as diffusion model activations and LLM KV cache, into a compact bitstream that can be transmitted over slower wires, such as Thunderbolt 3/4/5 networking between two consumer machines. The resulting 6x lossless compression on diffusion mid-block activations and 3x lossless on LLM KV cache provides significant bandwidth savings, making it particularly useful in scenarios where fast, low-latency data transfer is critical. ### Key Features: - **NVENC Compression Library**: A reusable codec library that efficiently compresses model state using dedicated NVENC silicon. - **Public Consumer Use**: The codec is used by the `comfyui-mesh` project to compress diffusion model activations, reducing the wall-clock time required for data transfer. - **6x Lossless on Diffusion Mid-Block Activations**: The codec achieves 6x lossless compression on diffusion model activations, significantly reducing the data transfer time. - **3x Lossless on LLM KV Cache**: The codec also provides 3x lossless compression on LLM KV cache, further optimizing data transfer efficiency. ### Tech Stack: - **Python**: The primary programming language used for implementing the codec library. ### Why It Matters: - **Bandwidth Savings**: The codec compresses model state into a compact bitstream, reducing the wall-clock time required for data transfer, especially over slower wires. - **Efficient Data Transfer**: By leveraging the idle NVENC silicon, the codec maximizes bandwidth efficiency,

Overview: This repository contains 514 Krea 2 Turbo prompts, each accompanied by an image generated by the Krea 2 Turbo model. These prompts are designed to be easily copied and used in various applications. Key features: - Easy copying of prompts and images - Downloadable all 514 prompts as a text file - Downloadable wildcard zip file for all category files - Available on ComfyUI and Hugging Face for integration - Included in ComfyUI workflows for easy integration Tech stack: HTML Why it matters: This repository provides a comprehensive set of Krea 2 Turbo prompts, making it a valuable resource for developers looking to incorporate AI-generated images into their projects.

Overview: This project is an implementation of the Jasper model, an end-to-end convolutional neural acoustic model, using PyTorch. It showcases the use of 1D convolutions, batch normalization, ReLU, dropout, and residual connections, demonstrating powerful performance. Key features: - Jasper 10x3 Model Test - Jasper 5x3 Model Test Tech stack: PyTorch Why it matters: This implementation provides a practical demonstration of a state-of-the-art model for speech recognition, showcasing the effectiveness of using only 1D convolutions and other standard neural network components. Repository: sooftware/jasper Primary language: Python GitHub stars: 32 GitHub description: PyTorch implementation of "Jasper: An End-to-End Convolutional Neural Acoustic Model" (INTERSPEECH 2019) GitHub license: Apache 2.0 GitHub build status: Passing Image: Jasper model architecture and usage example Usage: Example code for Jasper model training and usage Troubleshoots and Contributing: Issues for feedback, bug reports, and feature requests Code Style: Follows PEP-8 for documentation Reference: References to Jasper paper and NVIDIA DeepLearningExamples Author: Soohwan Kim [@sooftware]

Overview: SquareDev is a platform designed for developers to build applications powered by language models, offering features such as document loaders, vectors & embeddings, hosted large language models, and more. Key features: - Document Loaders (PDF, JSON, Website) - Vectors & Embeddings (Retrieval Augmented Generation, Semantic search) - Hosted Large Language Models (Open AI, HuggingFaceH4/zephyr-7b-beta) - Monitoring (coming soon) Tech stack: TypeScript Why it matters: SquareDev simplifies the process of building applications using language models, making it accessible to developers without AI expertise. Repository: squaredev-io/squaredev GitHub stars: 32 GitHub description: The platform for developers building applications powered by LLMs.