Projects

Overview: Detoxify is an AI-powered Chrome extension designed to filter out distracting content on YouTube, making it "impossible to be distracted" by focusing on your preferred categories of videos. Key features: - Dynamically filters your YouTube feed based on your preferred category - Pops up a dialogue box whenever you search/watch a video that doesn't align to your learning goal Tech stack: JavaScript Why it matters: This extension aims to enhance the user experience on YouTube by providing a cleaner, more focused browsing experience, making it easier to stay on task and achieve your learning goals. Repository: Saarthakkj/detoxify_yt Primary language: JavaScript GitHub stars: 11 GitHub description: Detoxify makes it "Impossible to Be Distracted on Youtube" .It Not only Dynamically Filters your Youtube based on your preferred category But also pops up a Dialogue Box whenever you search/watch a video that doesn't align to your learning goal.

### Overview: Sigma Studio is a modular AI platform designed to manage and orchestrate local and distributed AI workloads. It integrates with various hardware and software components, including GPUs, CPUs, and edge devices, to enable intelligent resource-aware model selection. ### Key Features: - 📥 **Model Management**: Seamlessly download and manage open-source models from repositories like Hugging Face and GGUF. - 📣 **Chat with Models**: Connect with local models via the SigmaEngine or external cloud providers for real-time token rendering. - 🧪 **Modular Workflow**: Switch between 20 predefined Modelfiles with a single click, enhancing productivity. - 🚀 **Orchestration**: Seamless integration across local and distributed hardware, ensuring efficient resource utilization. ### Tech Stack: Python ### Why it Matters: Sigma Studio transforms local hardware into a powerful AI development environment, making it easier to manage and orchestrate AI workloads efficiently. ### Repository: Sigmanih/SigmaStudio ### GitHub Stars: 11 ### Primary Language: Python ### GitHub Description: Sigma Studio is a modular AI platform for orchestrating local and distributed AI workloads. It manages LLMs, multimodal models, image generation, training, inference, and AI agents across GPUs, CPUs, and edge devices, with intelligent resource-aware model selection. ### Quick Start: 🚀 - 🔗 **Get Started**: Explore the repository and start using Sigma Studio. - 📝 **Documentation**: Follow the README for detailed instructions and tutorials. ### See Sigma Studio in Action:

### Overview: This project focuses on developing a CLIP-based multimodal representation learning framework for detecting hateful images on the Hateful Memes dataset and its inpainted/masked variants. The project utilizes prompt engineering (PE) to optimize LLM outputs and employs multiple LLMs for enhanced semantic alignment and multimodal representation learning. ### Key features: - Multimodal representation learning using CLIP and LLMs - Semantic alignment between image and text - Handling of image–text semantic inconsistency - Implicit references and sarcastic/ironic expressions - Optimized LLM output through prompt engineering ### Tech stack: Jupyter Notebook ### Why it matters: This project addresses critical challenges in hate speech detection, particularly in the context of social media, by leveraging advanced multimodal techniques to accurately identify and classify harmful content. The use of CLIP and LLMs ensures robust and effective detection, while prompt engineering enhances the accuracy and reliability of the model.

### Overview: TurboQuant-style KV-cache and vector compression for AI agents. ### Key features: - **KV-Cache Compression**: Efficiently compresses and decompresses KV-cache payloads to improve local and server-side inference performance. - **Vector Compression**: Compresses vector payloads to reduce storage and retrieval costs, making it easier for retrieval systems to store and rerank more data. - **Benchmarking**: Provides explicit benchmarking of quality, latency, and recall tradeoffs to help developers make informed decisions. ### Tech stack: Python ### Why it matters: `turboagents` is designed to complement existing AI systems by providing compression, reranking, and benchmarking capabilities without requiring a complete overhaul of the application. It integrates seamlessly with various AI frameworks, inference stacks, and retrieval systems, making it a valuable addition to the developer's toolkit.

Overview: This project is a Q&A system that leverages DistilBERT for text representation and Faiss for efficient similarity search in a vector database. It aims to provide accurate and relevant answers to user queries. Key features: - DistilBERT-based text representation - Faiss vector database for efficient similarity search - Streamlit web app interface Tech stack: Python Why it matters: This system efficiently retrieves relevant documents by leveraging pre-trained models and a vector database, making it suitable for applications requiring fast and accurate information retrieval.

### Overview: GLM-5.2 QuantTrio (Unpruned) on 4xDGX Spark (GB10): 327K-655K context or up to 5 concurrent agents, one recipe. TP4 + DCP + MTP ### Key Features: - Full model, no pruning - Holds to depth - One recipe, four lanes - Sane ops ### Tech Stack: Shell ### Why it Matters: - Deliberately newer stack than reference recipes were written against, tuned for current GB10 firmware. Repository: XanuNetworks/GLM-5.2-QuantTrio-DCP-4x-DGX-Spark Primary language: Shell GitHub stars: 11

Overview: QuotesMaker.py is a Python-based tool that converts quotes into images using pre-built templates. It supports converting single quotes or files containing multiple quotes, and offers various templates for customization. Key features: - Converts quotes into images using pre-built templates - Supports both manual input and file input - Offers 7 pre-built templates - Automatically selects random templates for multiple quotes - Accepts .png format for logos and supports .txt files for quotes Tech stack: Python Why it matters: This tool simplifies the process of creating visually appealing quotes, making it easier for users to personalize their quotes and share them with friends or on social media.

Overview: This repository showcases the integration of Red Hat OpenShift Data Science (Openshift AI) with various data science pipelines and model serving capabilities, including automated pipelines, streaming, and model registry. Key features: - Automated pipelines creation - Data science cluster deployment - Model registry setup - Kserve features for model serving Tech stack: Python, Docker, Kubernetes, OpenShift, Gitea, Grafana, MySQL Why it matters: This repository provides a comprehensive demonstration of how to leverage OpenShift AI for managing data science pipelines and deploying models, offering a practical approach to implementing MLOps in a cloud-native environment. Repository: adrien-legros/rhods-mnist GitHub stars: 11 GitHub description: Data science pipelines and model serving using Red Hat OpenShift Data Science

### Overview: Arthemy Live Tuner SDXL is a set of ComfyUI nodes designed to tweak SDXL, Illustrious, and NAI models by adjusting specific "slices" without training. It breaks down the U-Net and CLIP into areas of application, using sliders to boost or lower the intensity of any section. ### Key features: - Adjusts specific "slices" of SDXL, Illustrious, and NAI models. - Breaks down the U-Net and CLIP into areas of application (like Style, Lighting, Word Logic...). - Uses individual weights to boost or lower the intensity of any section without training. ### Tech stack: Python ### Why it matters: This tool offers granular control over a single model, transforming static checkpoints into a set of "Sliders" for precise control over the output. It's fast and can be seamlessly integrated into any workflow, making it a valuable addition to the developer's toolkit.

Overview: This project involves deploying a large language model (LLM) on AWS using only CPU for a question-answer chatbot application. The model is based on Llama2-7B with GGML quantization, and it leverages RAG (Retrieval Augmented Generation) to improve its performance with external data sources. Key features: - Deploying a large language model on AWS using only CPU - Leveraging RAG for improved performance with external data sources - Using Llama2-7B with GGML quantization for the LLM model Tech stack: Python Why it matters: This project showcases the effectiveness of using LLMs with RAG for improved accuracy and performance, especially when dealing with large datasets. It also highlights the importance of understanding and optimizing hyperparameters for optimal model performance.

Overview: Amortized is a project that aims to replace expensive frontier API calls with task models that are fine-tuned and optimized for specific tasks, such as classification, extraction, and summarization. These task models run on your infrastructure, cost significantly less, and don't break when the API provider changes. Key features: - Task models that replace frontier API calls - Fine-tuned models for specific tasks - Cost-effective and reliable task execution - Runs on your infrastructure, costs a fraction per inference - Small student models that match teacher models - Can be deployed using Kubernetes Tech stack: Python Why it matters: Amortized simplifies the process of building and deploying task models that can replace expensive API calls, reducing costs and ensuring reliability, making it a valuable tool for developers looking to optimize their infrastructure and reduce operational overhead.

### Overview: This repository provides a simple and efficient way to fine-tune an LLM (Large Language Model) hosted on Hugging Face using the `autotrain-advanced` method. The process involves converting Slack conversations into the required format, fine-tuning the model, and then merging the base model with the LoRA (Low-Rank Adaptation) to create a final output model. The fine-tuned model can be used for various applications, such as improving the performance of a specific user on a platform like Slack. ### Key Features: - Fine-tuning an LLM on your own hardware. - Utilizing the `autotrain-advanced` method from Hugging Face. - Converting Slack conversations into the OpenAI fine-tuning format. - Fine-tuning a base model pulled from Hugging Face (HF). - Merging the base model with the LoRA to create a final output model. - The process can be easily followed with a Quickstart example using Slack conversations. ### Tech Stack: The repository uses Python for the automation and conversion processes. ### Why it Matters: Fine-tuning an LLM on your own hardware can significantly reduce costs and improve the performance of the model. By fine-tuning the model on your own hardware, you can take advantage of your existing resources and avoid the high costs associated with cloud-based solutions. Additionally, fine-tuning the model on your own hardware can result in better performance and faster training times, as the model is optimized specifically for your hardware. ### Repository: The repository is hosted on GitHub and has been starred by

### Overview: This project involves the finetuning of the coqui-TTS model for voice cloning, utilizing a dataset of Priyanka Chopra's voice notes and transcriptions. The dataset is preprocessed and organized into a structured format, and the finetuned model is used to generate synthetic speech outputs. ### Key Features: - **Voice Cloning**: Converts text input into natural and expressive synthetic speech using the coqui-TTS model. - **Dataset Preparation**: The dataset consists of 212 voice notes and transcriptions, prepared through various steps including audio extraction, filtering, cleaning, and transcription. - **Model Training**: The coqui-TTS model is fine-tuned using specific hyperparameters for optimal performance. ### Tech Stack: - **Model**: coqui-TTS - **Dataset**: 212 voice notes and transcriptions - **Preprocessing Tools**: yt-dlp, Audacity, rnnoise - **Language**: Jupyter Notebook ### Why It Matters: This project demonstrates the effectiveness of using pre-trained models for text-to-speech conversion, showcasing the potential of voice cloning technology. The detailed dataset preparation and model fine-tuning ensure high-quality results, making it valuable for applications requiring precise voice replication.

### Overview: ArcMind Vector DB is a high-performance, flexible, and ergonomic vector similarity search database designed for the Internet Computer. It offers a general-purpose solution for AI-powered applications such as recommendation systems, search engines, and Retrieval Augmented Generation (RAG). ### Key features: - **Flexible Architecture**: Supports various AI models and applications. - **High Performance**: Optimized for fast vector similarity search. - **Scalability**: Can handle large-scale vector data beyond the canister storage limit. - **Embedding Model**: Uses the OpenAI text-embedding-ada-002 model for consistent results. - **Self-Hosted ML Models**: Integrates with ArcMind AI for long-term memory and scalable storage. ### Tech stack: Rust ### Why it matters: ArcMind Vector DB is crucial for applications requiring high-performance vector similarity search, making it an essential tool for AI-powered systems across various domains.

### Overview: MLX-KQuant is a C++ extension for the MLX framework, providing native K-quant support on Apple Silicon. It includes a set of K-quant superblocks and per-block integer codecs, as well as a toolchain for quantizing models into K-quant MLX safetensors checkpoints and fine-tuning them. ### Key Features: - **Ops (C++/Metal)**: A `kq.*` namespace with ten K-quant/legacy codecs, plus nine IQ codecs, and a set of DeepSeek/GLM sparse-attention kernels. - **Tooling (Python)**: A CLI for quantization, fine-tuning, LoRA training, and fusion, including a `loader` for MLX-native safetensors checkpoints. ### Tech Stack: - **C++/Metal**: The MLX extension with custom Metal kernels for K-quant operations. - **Python**: A CLI for quantization, fine-tuning, LoRA training, and fusion. ### Why It Matters: K-quants have roughly half the divergence compared to affine quants at the same bitrate, making them a more efficient alternative for MLX models. The toolchain provided by MLX-KQuant ensures that models can be quantized and fine-tuned seamlessly on Apple Silicon, offering significant performance benefits.

Overview: Local AI is a fully offline, portable workspace for downloading, managing, and chatting with powerful Ollama models. It features smart context memory, seamless model switching, and real-time system monitoring. Key features: - Fully offline & private - Smart context memory - Portable workspace - Live model discovery - Vision & image support - Seamless model switching - Real-time system monitor - GPU-accelerated image generation - Local network server Tech stack: Python Why it matters: Local AI provides a secure, efficient, and user-friendly way to interact with AI models, eliminating the need for cloud services and ensuring privacy and performance. Repository: ashser004/Ollama-UI Primary language: Python GitHub stars: 11 GitHub description: Local AI is a fully offline, portable workspace where users can download, manage, and chat with powerful Ollama models privately, featuring smart context memory.

Overview: This project provides a comprehensive guide on fine-tuning BGE embeddings using synthetic data from Amazon Bedrock, aimed at developers looking to enhance their machine learning skills. Key features: - Fine-tuning BGE embeddings - Generating synthetic data from Amazon Bedrock - Running the code in a Jupyter notebook environment - Deploying the fine-tuned BGE model using Amazon SageMaker Tech stack: Jupyter Notebook Why it matters: This project serves as a practical tutorial for developers interested in leveraging Amazon Bedrock and Amazon SageMaker for advanced natural language processing tasks, specifically focusing on BGE embeddings fine-tuning.

Overview: Seedream 5.0 Pro API provides a comprehensive set of image and video generation capabilities with precise text rendering and multi-reference editing, all at a flexible and efficient price. Key features: - **Model ID and Routes**: Seedream 5.0 Pro (`seedream-5-0-pro`) offers 1K/2K image models with layered pricing, precise text rendering, and multi-reference editing. - **Pricing**: Prices are per delivered image, with tiered pricing based on resolution, ranging from 0.014625 per image for 1K to 0.0365625 per image for 1K-layered models. - **Request Parameters**: Accepts prompt, size, resolution, number of images, and reference images for editing tasks. Tech stack: Python Why it matters: Seedream 5.0 Pro leverages the OpenAI-compatible APIMart endpoint, providing a single key for seamless access to Seedream 5.0 Pro models across various resolutions and features, eliminating the need for multiple API keys and unnecessary costs. Repository: ayekoa913g/seedream-5-0-pro-api-id Primary language: Python GitHub stars: 11 GitHub description: Seedream 5.0 Pro API (`seedream-5-0-pro`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Seedream 4.0 Prompts: A collection of 8 tested recipes for high-quality image prompts, including aspect ratios, model ID, and pricing details. Key features: - Model: `seedream-4-0` - Recipes: 8 (machine-readable in `data/prompts.json`, table in `PROMPTS.md`) - Effective price used: $0.0195 per image — estimated cost of the whole set: $0.156 Tech stack: Python Why it matters: This resource provides a structured way to create high-quality image prompts using Seedream 4.0, offering precise control over the subject, light, framing, and setting, all while keeping costs under $0.10 per image. Repository: b2n87l0aoum6e/seedream-4-0-prompts-de Primary language: Python GitHub stars: 11 GitHub description: Seedream 4.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedream-4-0`) and the cost per image. README excerpt: # Seedream 4.0 Prompts — 8 Tested Recipes With Outputs <!-- conv-kit:v1 --> <p align="center"> <img src="assets/badges/price.svg" alt="observed unit price"> <img src="assets/badges/billing.svg" alt="billing model"> <img src="assets/badges/compat.svg" alt="OpenAI-compatible endpoint">

### Overview: Seedance 1.5 Pro Prompts: 8 Tested Recipes with Outputs Dreamina — Seedance 1.5 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-1.5-pro`) and the cost per image. This repository includes 8 recipes, each with an example render, duration, and cost on the APIMart route. ### Key Features: - **Model ID:** `seedance-1.5-pro` - **Cost per Image:** $0.0104 per second - **Duration:** 5 seconds - **Prompt Structure:** - **Subject:** Concrete object or person with material - **Light:** Direction + quality - **Framing:** Ratio + lens or distance - **Setting:** Surface, backdrop, environment - **Finish:** Render or film character ### Tech Stack: Python ### Why It Matters: - **Access to 300+ Models:** Seedance 1.5 Pro offers access to over 300 image, video, and language models. - **Cost Per Second:** $0.0104 per second, with an estimated cost of $0.0416 for the entire set. - **Real-Time Pricing:** The cost is reflected in the response, and tasks marked as failed are refunded. - **Asynchronous Nature:** Tasks are batched and retried in the background, not integrated as part of a real-time service. ### Repository: b