Projects

Overview: Scubiee is a local code-context engine designed to save up to 30% of AI coding tokens by providing semantic and structural code search capabilities. It integrates with various AI coding agents and supports live sync and hardware acceleration. Key features: - Semantic and structural code search through `map` (`find` / `focus`) and terminal command `scubiee search`. - Fused retrieval over a code graph, BM25 lexical match, and dense embeddings stored in FAISS. - Tree-sitter parsing for Python, JavaScript, TypeScript, Go, Rust, Java, C, C++, C#, Ruby, and JSON. - Live sync: background daemon diffs the repo and re-indexes only what changed. - One-command MCP setup for Cursor, Claude Code, Codex, Copilot, Kiro, Windsurf, Cline, Roo Code, Continue, Zed, OpenCode, Amp, Devin Desktop, and Pi (`scubiee connect --<tool>` or `--all`). - Hardware acceleration picked automatically at setup: CUDA, DirectML, Apple MLX, or CPU. - Runs on Windows, macOS, and Linux with Python 3.11+. Tech stack: Python 3.11+ with hardware acceleration. Why it matters: Scubiee offers a local, always-fresh index for AI coding agents, reducing the need for cloud resources and improving efficiency in AI coding workflows. Repository: Usmansayed/scubiee

Overview: Grok Image API is a powerful image generation service that offers text-to-image and edit functionalities, including five aspect ratios and up to 10 images per request. It supports OpenAI-compatible endpoints and is priced on a per-image basis, with a minimum charge of $0.015 per image, making it ideal for teams looking to integrate Grok Imagine 1.5 into their applications without the need for a monthly subscription. Key features: - Supports text-to-image and reference-based editing - Offers five aspect ratios and up to 10 images per request - Per-image pricing with a minimum of $0.015 per image - Compatible with OpenAI-compatible endpoints Tech stack: Python Why it matters: Grok Image API is a valuable tool for developers looking to integrate advanced image generation capabilities into their applications without the need for a monthly subscription. It simplifies the process of using multiple models and provides a straightforward API interface, making it easy to integrate into existing projects. GitHub stars: 145 Repository: apimart-API-Gateway/grok-image-api

### Overview: Next-token prediction is a JavaScript library designed to create high-quality language models based on a body of text. It provides tools for predicting the next word, phrase, or pixel, making it useful for applications like autocomplete, autocorrect, spell checking, and more. ### Key features: - Predict the next word - Predict the next phrase - Predict the next pixel - Complete phrases - Get top k sample predictions ### Tech stack: Next-token prediction is built using JavaScript. ### Why it matters: This library simplifies the process of creating language models, making it easier to build fast and accurate language and diffusion models.

Overview: This project showcases various BentoML models and their applications, providing a visual and interactive experience. Key features: - Model visualization - Model deployment - Model explanation - Model comparison Tech stack: Python, BentoML, Dash, Plotly Why it matters: It demonstrates the capabilities and usability of BentoML, making it easier for developers to create and deploy machine learning models.

Overview: This repository contains the code for the paper "DiffUTE: Universal Text Editing Diffusion Model" presented at NeurIPS 2023. It includes the implementation of a universal text editing diffusion model using the AnyText dataset. Key features: - Universal text editing diffusion model - Reproduces method using the AnyText dataset - Available weights for download and demo Tech stack: Python Why it matters: This repository provides a practical implementation of the DiffUTE model, showcasing its capabilities and utility in text editing tasks.

Overview: TruthX is an inference-time method designed to enhance the truthfulness of large language models (LLMs) by editing their internal representations in a truthful space. This approach aims to mitigate hallucinations in LLM responses. Key features: - TruthX can control LLMs to generate both truthful and hallucinatory responses. - It is compatible with various LLM models, including Llama-2-7B-Chat, Alpaca-7B, Llama-2-7B, Llama-2-13B-Chat, Vicuna-7B-v1.5, Mistral-7B-v0.1, Mistral-7B-Instruct-v0.1, and Mistral-7B-Instruct-v0.2. Tech stack: Python and PyTorch. Why it matters: TruthX significantly improves the truthfulness of responses generated by LLMs on benchmarks like TruthfulQA, achieving an average enhancement of 20% across 13 advanced LLMs. This makes it a valuable tool for researchers and developers looking to enhance the reliability of AI systems.

Overview: This project is a web-based interface extension for ComfyUI, designed to provide a stable and user-friendly experience for inference-focused workflows. It includes features such as dynamic state preservation, advanced queue management, prompt presets, batch support, and a minimalist image editor. Key features: 1. Stable UI states 2. Compatibility with ComfyUI workflows 3. Advanced queue management 4. Prompt presets 5. Batch support 6. Built-in minimalist image editor Tech stack: Python Why it matters: This project simplifies the workflow of ComfyUI users by providing a stable and intuitive interface, allowing them to focus on inference without the complexity of traditional UIs. It also offers a more modern and user-friendly experience compared to existing solutions like SwarmUI or ViewComfy. Repository: light-and-ray/Minimalistic-Comfy-Wrapper-WebUI Primary language: Python GitHub stars: 146 GitHub description: MCWW: Additional non-node based UI for ComfyUI focused on inference. Stable UI states; presets; and advanced queue. Based on Gradio

Overview: This Blazor Web App and Minimal API enables performing RAG and vector search using Azure SQL Database and Azure OpenAI, leveraging the native `VECTOR` type for efficient semantic search and document embedding. Key features: - Load documents (PDF, DOCX, TXT, MD) - Generate embeddings and save them as vectors in Azure SQL Database or SQL Server 2025 - Perform semantic search and RAG using Azure OpenAI and Microsoft Agent Framework agents - Interact via a Blazor Web App or programmatically via Minimal API Tech stack: C# Why it matters: This application provides a comprehensive solution for integrating vector search and RAG capabilities into a Blazor application, utilizing Azure SQL Database and Azure OpenAI services, making it suitable for applications requiring advanced semantic search and AI-driven content management.

### Overview: PyRobBot is a Python package that uses OpenAI's large language models (LLMs) to create a fully configurable personal assistant. It supports voice chat, internet access, and a web browser interface, along with a text-based chat interface. The assistant can search the web, provide news and weather updates, and integrate with a Streamlit app for enhanced user experience. ### Key Features: - **Voice Chat**: Continuous voice input and output, no need to press buttons. - **Internet Access**: Searches the web for answers not found in training data. - **Web Browser Interface**: Streamlit app for a more traditional chatbot experience. - **Text Interface**: Text-based chat with voice prompts and replies. - **Terminal Chat**: A more "Wake up, Neo" experience. - **Configurability**: Supports multiple languages, multiple LLMs, and customizable parameters. ### Tech Stack: Python ### Why It Matters: PyRobBot leverages OpenAI's powerful language models to create a versatile and customizable personal assistant, making it a valuable tool for anyone looking to enhance their communication and information-seeking experience.

### Overview: InstantRestore is an innovative single-step face restoration framework designed for real-time processing and identity preservation. Leveraging shared-image attention and a landmark attention loss, it outperforms existing methods in both quality and speed, making it an ideal choice for large-scale applications. ### Key Features: - **Single-Step Restoration**: InstantRestore achieves near-real-time performance by performing a single forward pass through the network. - **Shared-Image Attention**: Utilizes shared image attention to efficiently preserve the input identity. - **Landmark Attention Loss**: Aligns key facial landmarks to refine the attention maps, enhancing identity preservation. - **Fast and Scalable**: Suitable for large-scale applications due to its scalable solution and fast inference times. ### Tech Stack: Python ### Why It Matters: InstantRestore offers a fast, efficient, and identity-preserving face restoration solution that outperforms existing methods in both quality and speed, making it an appealing choice for large-scale applications.

### Overview: Moe-L2 is an advanced model offloading system designed for low-VRAM GPUs, enabling the efficient execution of 100B+ models like DeepSeek, Qwen, and Mixtral on 10-11 GB cards. It utilizes an expert cache with LRU hot experts and an OpenAI-compatible proxy, significantly reducing VRAM usage while maintaining performance. ### Key features: - **Expert Cache with LRU Hot Experts**: Preloads frequently used experts into a shared-memory LRU cache for quick access. - **Transparent Proxy**: Predicts which experts are needed for prompts and preloads them into the cache. - **GPU-Optimized**: Designed to work seamlessly with 10-11 GB GPUs, providing up to 60% VRAM savings. - **Selective Pinning**: Optimizes memory usage by pinning experts on-demand, reducing VRAM usage without sacrificing performance. ### Tech stack: Cuda ### Why it matters: Moe-L2 allows developers to run high-performance models on GPUs with limited VRAM, making it ideal for applications requiring large models without the need for excessive GPU resources. This system ensures that models can be executed efficiently on lower-end GPUs, enabling more cost-effective and resource-efficient computing solutions.

Overview: This comprehensive guide provides a mental-model-first approach to understanding and working with Audio AI, covering foundational concepts, ASR, TTS, audio classification, and more. It includes bilingual (English/Farsi) content and focuses on decision-making and practical applications rather than just theoretical explanations. Key features: - Mental-model-first approach - Bilingual (English/Farsi) content - Covers ASR, TTS, audio classification, Transformer architectures, and real-world pipelines like voice assistants and meeting transcription - Includes a table of contents with sections on foundations, ASR, TTS, and systems, cross-cutting comparison tables, glossary, and key takeaways Tech stack: None specified Why it matters: By focusing on the decision-making process rather than just the technical details, this guide helps developers understand when and why certain architectures and techniques should be used, making it a valuable resource for practical application in audio AI projects.

Overview: FATE-Serving is a scalable, high-performance serving system specifically designed for federated learning models, optimized for production environments. Key features: - High performance online Federated Learning algorithms. - Real-time inference using federated learning models. - Support for parallel inference between guest and host. - Parallel computing in inference requests. - Service management through Zookeeper registry. - Visualization tools for cluster and model management. Tech stack: Java Why it matters: FATE-Serving is crucial for deploying federated learning models in production, ensuring efficient and scalable model serving with real-time inference capabilities.

Overview: MONAI Deploy App SDK provides a framework and tools for designing, developing, and verifying AI-driven medical imaging applications in the healthcare domain. Key features: - Flexible Python API for building inference applications - Easy management of inference applications via DAGs - Built-in operators for DICOM data ingestion - In-proc PyTorch inference and remote Triton Inference Server support - Built-in MONAI pre and post transformations - Package inference applications into portable MONAI Application Packages - Locally run and debug inference applications using App Runner Tech stack: Python Why it matters: This SDK simplifies the process of developing AI-driven medical imaging applications, making it easier to design, test, and deploy applications in the healthcare domain.

Overview: This project is a continuous speech recognizer implemented in Kotlin that listens for a specific activation keyword and triggers speech recognition. It is designed to be used with care to avoid battery drain. Key features: - Continuous listening - Activation keyword detection - Sound system for activation - Custom callback interface for keyword detection and speech recognition availability Tech stack: Kotlin Why it matters: This project is useful for developers who need a continuous speech recognition system that can be triggered by a specific keyword, making it ideal for applications requiring voice commands or automated responses. Repository: StephenVinouze/KontinuousSpeechRecognizer GitHub stars: 144

Overview: This project enhances the construction of RAG applications by addressing challenges, improving accessibility, scalability, and managing data and user access. It uses Django, Llamaindex, and Google Drive for effective database management. Key features: - Managing Data Sources for Google Drive with Scheduler - User Management with Access Control and Permission Settings - RAG API - Admin panel Tech stack: Python, Django, Llamaindex, Google Drive Why it matters: This project simplifies the process of building RAG applications by providing a comprehensive solution that addresses common challenges, enhances accessibility, and optimizes scalability. It integrates modern tools like Llamaindex and Google Drive, making it an efficient and scalable solution for RAG applications.

### Overview: Arkor is a TypeScript framework designed for fine-tuning open-weight language models, providing a type-safe configuration, local Studio, and managed GPUs for efficient development. ### Key features: - **Fine-tuning from a TypeScript file.** `createTrainer({ model, dataset, lora, ... })` configured LoRA training against a supported base model. - **Three curated templates.** `triage`, `translate`, and `redaction` paired a Gemma 4 base model with public HuggingFace datasets. - **Typed lifecycle callbacks.** `onStarted`, `onLog`, `onCheckpoint`, `onCompleted`, and `onFailed` exposed training events to application code. - **Checkpoint inference.** Inside `onCheckpoint`, `infer({ messages })` could test a checkpoint while training was still running. - **Local Studio.** `arkor dev` opened a web UI with a jobs list, loss chart, logs, and a Playground. - **Anonymous workspaces and OAuth.** The original onboarding flow supported anonymous workspaces that could later be linked with OAuth. ### Tech stack: TypeScript ### Why it matters: Arkor simplifies the fine-tuning process for developers, making it easier to integrate models into existing TypeScript projects and workflows. ### Repository: arkorlab/arkor ### GitHub stars: 144 ### GitHub description: The Best TypeScript framework for fine-tuning open-weight LLMs.

Overview: llm.rb is a runtime designed for building agentic AI applications on Ruby, providing a cohesive API across multiple providers and a single, coherent tool loop. Key features: - Supports concurrent and parallel tool execution - Single coherent API spanning multiple providers - Handles exceptions and tool call failures gracefully - Runs on Ruby's standard library by default Tech stack: Ruby Why it matters: llm.rb simplifies the process of integrating AI capabilities into Ruby applications by providing a unified interface and robust error handling, making it easier to build and deploy agentic AI solutions on Ruby-based platforms.

Overview: This repository contains a collection of papers on augmented Large Language Models (LLMs) with self-reflection capabilities. These models are designed to improve their performance by incorporating self-correction strategies, self-improvement techniques, and self-verification mechanisms. Key features: - Papers on various self-correction and self-improvement strategies - Contributions to the field of language models through self-reflection - Techniques for enhancing the accuracy and reliability of LLMs Tech stack: Python, LaTeX, Markdown, and various programming languages for implementing and documenting the papers. Why it matters: The repository highlights the advancements in the field of language models that incorporate self-reflection, aiming to make these models more robust and reliable through continuous learning and self-correction.

Overview: RALF is an official repository for the CVPR 2024 Oral paper "Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation". This repository provides a comprehensive package for content-aware layout generation, including state-of-the-art methods and detailed documentation. Key features: - RALF: Retrieval-augmented content-aware layout generation - Autoreg Baseline: State-of-the-art method for content-aware layout generation - Dataset splits, pre-processing, training, and inference methods - Benchmark comparison with other state-of-the-art methods Tech stack: Python Why it matters: RALF offers a practical solution for content-aware layout generation, making it easier to implement and evaluate various layout generation techniques. Repository: CyberAgentAILab/RALF GitHub stars: 143