Projects

Overview: Qwisp is a fast, optionally-lossless local inference engine for Qwen3.6-35B-A3B on Apple Silicon, serving over an OpenAI-compatible HTTP API. Key features: - **Runtime = Tell**, **engine = Seedless** (raw Metal; MLX is the substrate). - **Lossless option**: the strict path reproduces the quantised greedy token stream bit-for-bit. - **Drop-in**: OpenAI `/v1/chat/completions` (SSE) + `/v1/models`, plus a `qwisp chat` CLI. Tech stack: Swift Why it matters: Qwisp is designed for efficient inference on Apple Silicon, leveraging raw Metal for faster decoding, while maintaining compatibility with OpenAI's API and serving as a drop-in replacement for OpenAI models. Repository: penta2himajin/qwisp GitHub stars: 17

Overview: VSB is a benchmarking suite for vector search, designed to measure the performance of various vector databases. Key features: - Supports multiple databases including Pinecone, pgvector, and OpenSearch - Runs workloads like MNIST and NQ768 against these databases - Measures throughput, latency, and other metrics - Automatically runs setup, populate, and run phases - Writes detailed results to stats.json and displays a summary Tech stack: Python Why it matters: VSB provides a quick and efficient way to benchmark the performance of vector databases, helping developers and researchers optimize their search solutions.

Overview: This project implements a semantic retrieval recommender and information retrieval system based on Google Cloud infrastructure, built on the Google Cloud Platform (GCP). The system leverages machine learning techniques to enhance the shopping experience by providing relevant product suggestions and personalized content. Key features: - Semantic product search based on natural language queries - Extraction of key product features - Sentiment analysis of product reviews - Contextual recommendations Tech stack: TypeScript Why it matters: This project offers a smarter way to shop online by integrating machine learning techniques with Google Cloud infrastructure, providing enhanced product recommendations and personalized content. GitHub stars: 17 Repository: polarbear333/rag-llm-based-recommender

Overview: OpenManga is a self-hosted production tool that converts stories into consistent AI-illustrated comics, webtoons, or narrated videos. It allows users to bring their own provider keys, local TTS, deterministic lettering, and exports. Key features: - Your own provider keys (BYOK) for text reasoning and raster generation - Kokoro-82M for local speech synthesis - The application handles everything else - No telemetry, no analytics, no phone-home - Production run: one button to run the entire pipeline - Update production: runs only the steps that need to be updated - Revision: shows what is out of date from story to video Tech stack: TypeScript Why it matters: OpenManga is a self-hosted solution for consistent AI-generated content, making it ideal for small to medium-sized projects that require high-quality illustration and narration. Repository: pr0h0/OpenManga GitHub stars: 17

Overview: This project introduces an OpenClaw memory plugin backed by a self-hosted Mem0 REST API, replacing the default LanceDB backend with Mem0's semantic extraction pipeline. The plugin enhances agent recall and memory management capabilities, offering features like auto-recall, auto-capture, and a zero-native-dependency setup. Key features: - Three agent tools: `memory_recall`, `memory_store`, `memory_forget` - Auto-recall: Injects relevant memories before each agent execution - Auto-capture: Stores conversation highlights after each agent execution - CLI commands: `openclaw mem0 search|list|forget` - Graceful degradation: Logs warnings if the Mem0 server is unreachable, never crashes the gateway - Zero native dependencies: Uses Node.js built-in `fetch()` — no SDK required Tech stack: Node.js with TypeScript Why it matters: This plugin leverages Mem0's semantic extraction pipeline, enhancing the long-term memory capabilities of OpenClaw agents, making them more versatile and powerful.

### Overview: Valori is a high-performance vector database library for Python that offers efficient storage, indexing, and search capabilities for high-dimensional vectors. ### Key features: - **🚀 High Performance**: Optimized for speed with multiple indexing algorithms. - **📄 Document Parsing**: Support for PDF, Office, text, and advanced parsing with Docling. - **🔄 Processing Pipeline**: Complete document processing with cleaning, chunking, and embedding. - **💾 Multiple Storage Backends**: Memory, disk, and hybrid storage options. - **🔍 Advanced Indexing**: Flat, HNSW, and IVF indices for different use cases. - **🗜️ Vector Quantization**: Scalar and product quantization for memory efficiency. - **💾 Persistence**: Tensor-based and incremental persistence strategies. - **🏭 Production Ready**: Comprehensive logging, monitoring, and error handling. - **🐍 Python Native**: Pure Python implementation with NumPy integration. - **📊 Extensible**: Plugin architecture for custom components. ### Tech stack: Python ### Why it matters: Valori is a powerful tool for developers looking to efficiently store, index, and search high-dimensional vectors in Python. Its high performance, versatility, and extensibility make it suitable for a wide range of applications, from natural language processing to image retrieval.

### Overview: VideoDB Chat is a frontend interface designed for building chat-based systems and connecting with agent-driven workflows. It integrates seamlessly with Director, a powerful tool for creating complex workflows and chat interfaces. ### Key Features: 1. **Chat Interface Components**: Provides essential chat UI components for integrating with Director. 2. **Customizable Chat Hook**: Allows developers to configure their own chat hook for better integration with Director. 3. **Real-Time Updates**: Reactively updates conversations in real-time, ensuring the chat interface remains current. 4. **Responsive Design**: Designed to be easily integrated into various frontend frameworks, including Vue. ### Tech Stack: Vue ### Why It Matters: VideoDB Chat simplifies the process of integrating chat functionality with Director, making it easier for developers to create robust chat-based workflows. Its real-time updates and customizable chat hook options make it a valuable tool for teams looking to enhance their chat interfaces. ### Repository: Repository: [video-db/videodb-chat](https://github.com/video-db/videodb-chat) GitHub Stars: 17 ### Primary Language: Vue ### GitHub Description: Frontend interface for building chat based system and connecting with agent driven workflows. ### GitHub Description (excerpt): [![NPM version][npm-shield]][npm-url] [![Stargazers][stars-shield]][stars-url] [![Issues][issues-shield]][issues-url] [![Website][website-shield]][website-url] [![Discord][discord-shield]][discord-url] <br /> <p align="center"> <a href="https://

Overview: This project provides a Python SDK for monitoring and tracking performance metrics of local and remote ML inference using the WildEdge platform. It tracks latency, token usage, errors, model metadata, and local hardware context across various libraries like llama.cpp, MLX, Transformers, OpenAI, and more. Key features: - Tracks performance metrics of local and remote ML inference - Instrument supported libraries with one command or in your code - Does not upload prompts or outputs unless explicitly attached - Supports multiple models and integrations Tech stack: Python Why it matters: This SDK simplifies the monitoring and tracking of ML inference performance, making it easier to analyze and optimize the performance of various models and libraries used in AI and ML applications. Repository: wild-edge/wildedge-python Primary language: Python GitHub stars: 17 GitHub description: Python SDK for WildEdge README excerpt: Monitor local and remote ML inference from Python. Track latency, token usage, errors, model metadata, and local hardware context across various libraries like llama.cpp, MLX, Transformers, OpenAI, and more. Instrument supported libraries with one command, or add tracking in your code. Prompts and outputs aren't uploaded unless you explicitly attach them. [Get started](#install) · [Explore the demo](https://wildedge.dev/demo) · [Docs](https://docs.wildedge.dev) · [Website](https://wildedge.dev) | Model performance | Workflow trace | SQL editor | |:---:|:---:|:---:| | <a href="

### Overview: NexcoreMultimodalSkill is a comprehensive Agent Skill that combines OCR, CNN, VLM, and Agnes media generation into a single agent, offering visual evidence, image editing, video generation, and local image creation. ### Key features: - **Primary visual evidence**: NexcoreVision runs local OCR, optional local CNN inference, and an independent vision model. - **Fallback and document parsing**: `vision.py` parses PDFs and Office files or continues with browser and API vision engines. - **Media output**: `agnes_api.py` generates images, edits images, generates videos, and queries asynchronous video tasks through Agnes AI. - **Local image generation**: `comfyui_local.py` generates or edits images with Qwen-Image-2.1 on the user's own ComfyUI installation. ### Tech stack: Python ### Why it matters: NexcoreMultimodalSkill provides a unified interface for text-only agents, offering visual evidence, image editing, video generation, and local image creation, making it a versatile tool for evidence analysis and media processing.

### Overview: This project is a Python script designed to create a weighted, category-balanced dataset for fine-tuning Large Language Models (LLMs). It takes a directory of JSONL files containing OpenAI chat-format conversations, samples from each category according to configurable weights, and outputs a single shuffled dataset. ### Key Features: - **Parallel File Loading**: Reads all JSONL files concurrently using a thread pool. - **Weighted Sampling**: Each category gets `total_examples * weight` rows. Oversamples with replacement if the source data is smaller than the target. - **Category Extraction**: Uses the first assistant message's content as the category label. - **Shuffled Output**: Final dataset is shuffled and written as JSONL. - **Evaluation Script**: Inspects the output, counts category distribution, and renders a terminal table. ### Tech Stack: - Python ### Why It Matters: This script is essential for preparing datasets for fine-tuning language models, ensuring that each category contributes a controlled proportion to the final dataset. It helps in balancing the dataset and improving the model's performance across different categories. ### Repository: - **Repository**: yigitkonur/cli-finetune-dataset - **GitHub Stars**: 17 - **GitHub Description**: weighted category-balanced dataset builder for LLM fine-tuning - **Primary Language**: Python ### Install: ```bash pip install pandas rich ``` ### Configure: Edit `config.ini`: ```ini [Paths] jsonl_directory = /path/to/your/jsonl/files output_file = /path/to

### Overview: LookAhead Tuning is a project that introduces two data-centric methods to enhance the safety of language models through partial answer previews. This approach reduces the loss associated with answer prefix tokens, ensuring the model remains safe during training. ### Key Features: - **Real Mode**: Uses a tokenizer to concatenate a specified number of tokens from the output field to the input field. - **Virtual Mode**: Inserts a predefined string into both the input and output fields without tokenization. - **Customizable Fields**: Allows specification of which fields in JSON objects to process. - **Easy Configuration**: Simple command-line arguments for tailoring the processing to specific needs. ### Tech Stack: Python ### Why it Matters: LookAhead Tuning aims to improve the safety of language models by minimizing disturbances to the model's initial tokens, ensuring the model remains safe during training.

Overview: llmon-py is a work-in-progress (WIP) multimodal web interface designed for real-time inference of Llama 3-8B models using text, voice, and function calls. Key features: - Real-time text-based inference - Voice-based inference - Function-based inference - User-friendly interface Tech stack: Python Why it matters: llmon-py provides a convenient and efficient way to interact with large language models, leveraging multiple modalities to enhance user experience and flexibility.

Overview: LLM-Emu is a profile-driven online emulator designed to replicate the behavior of vLLM without the need for a GPU, allowing developers to benchmark and test their models in a more controlled environment. Key features: - LLM-Emu plugin (~1.7K LoC online + ~0.6K LoC offline tooling) - Oracle for real-GPU profile capture (adaptive-K Shepard pooling) - Intercepts execute_model() to return a pending Future with predicted step latency - Mocks torch.cuda for CPU-only host mode - Supports vLLM v0.18.1 with vllm_patches Tech stack: Python Why it matters: LLM-Emu simplifies the process of emulating LLM serving systems, enabling developers to test and benchmark their models in a more controlled and efficient environment without the need for a GPU, thus reducing the complexity and cost of the testing process.

Overview: Carter is a secure, privacy-focused platform designed to help users manage and share links in an organized and controlled manner. Key features: - Secure interface - Link categorization and retrieval - Full control over link access Tech stack: TypeScript Why it matters: Carter empowers users with full control over their links, ensuring privacy and accessibility. Repository: Alkush-Pipania/Carter Primary language: TypeScript GitHub stars: 16

Overview: This project is a code repository for the paper "Learning Adaptive Fusion Bank for Multi-modal Salient Object Detection," which was accepted at TCSVT 2024. The repository includes the prediction results and weights based on VGG and ResNet backbones, as well as the citation information for the work. Key features: - Prediction results and weights based on VGG and ResNet backbones - Citing information for the paper - Framework diagram - Performance comparison with RGB-D and RGB-T SOD datasets - Pretrained models and their parameters Tech stack: Python Why it matters: This project is significant for researchers and developers interested in multi-modal salient object detection and adaptive fusion techniques. Repository: Angknpng/LAFB GitHub stars: 16 GitHub description: Learning Adaptive Fusion Bank for Multi-modal Salient Object Detection, accepted at TCSVT 2024.

### Overview: This repository is a community-maintained catalog of open and low-filter AI image models, checkpoints, and local inference resources. It focuses on filtering and practical trade-offs, providing a comprehensive comparison of various uncensored AI image models. ### Key features: - Local open-weight models - Community-reported uncensored or low-filter hosted endpoints - Community checkpoints and fine-tunes - Established local baselines ### Tech stack: Community-maintained catalog ### Why it matters: This catalog helps users find uncensored or low-filter AI image models that are not guaranteed to be uncensored, providing a detailed comparison of models across different access methods, providers, and versions. ### GitHub stars: 16 ### Repository: Anil-matcha/awesome-uncensored-ai-image-models

### Overview RiverSnap is a local citizen science project focused on estimating hydraulic parameters of rivers using machine learning and AI models. The project aims to develop and investigate novel methods to determine critical indices related to river systems, leveraging cutting-edge techniques for accurate water segmentation from close-range remote sensing imagery. ### Key Features - **Data Collection**: Records and estimates water level, river course, and vegetation changes using close-range images. - **Methodology Focus**: Emphasizes the methodology over data collection, with a focus on developing and investigating novel methods. - **Techniques**: Utilizes machine learning and AI models, particularly the Segment Anything Model (SAM), to accurately determine hydraulic parameters. ### Tech Stack - Python ### Why It Matters RiverSnap is crucial for advancing the field of river hydraulic parameter estimation by providing a robust, accurate method for analyzing water-related parameters in real-time. This project not only contributes to the scientific understanding of river systems but also has practical applications in water management and disaster risk assessment. ### Repository - Repository: ArminMoghimi/RiverSnap - GitHub Stars: 16 - GitHub Description: RiverSnap - estimation of river hydraulic parameters using machine learning/AI models ### Primary Language - Primary Language: Jupyter Notebook ### GitHub Description RiverSnap is a local citizen science project that records and estimates some hydraulic parameters such as water level (water level), river course (erosion/subsidence), and vegetation changes around the river bank using close-range images shared by citizens or other ways. The project RiverSnap is focused on further developing and

Overview: TySVA is a TypeScript Voice Assistant that leverages Qdrant for local database, LinkUp for web deep search, MCP servers for vector search and web search automation, ElevenLabs for voice input transcription and voice output generation, and LlamaIndex for agent workflows. It supports voice input/output and textual input/output. Key features: - Integrates Qdrant for TypeScript documentation - Utilizes LinkUp for comprehensive web search - Employs MCP servers for vector and web search automation - Leverages ElevenLabs for voice input transcription and output generation - Supports LlamaIndex for agent workflows Tech stack: Python Why it matters: TySVA provides a user-friendly environment for learning TypeScript by integrating powerful tools for documentation, search, and voice interaction, making it an excellent resource for developers looking to enhance their TypeScript skills.

### Overview: Fractured Chain-of-Thought Reasoning (frac-cot) is an efficient 3D sampling method designed for long-CoT LLMs, focusing on improving pass@k scores across multiple dimensions. This unified inference-time strategy interpolates between full CoT and solution-only sampling along three orthogonal axes: reasoning trajectories, final solutions, and reasoning depth. ### Key Features: - **Unified Inference-Time Strategy**: Interpolates between full CoT and solution-only sampling along three orthogonal axes. - **Efficient Sampling**: Offers the highest pass@k by sampling over all three dimensions. - **Process Reward Model**: Selects the best solution among all 3D samplings, achieving a ~10% accuracy improvement. - **Self-Consistency Property**: Early stops generation, saving 20% tokens without sacrificing accuracy. ### Tech Stack: - **Python**: The primary programming language used for development. ### Why It Matters: Fractured Chain-of-Thought Reasoning significantly improves the pass@k scores across multiple dimensions, making it a valuable tool for enhancing the performance of long-CoT LLMs.

### Overview: LoVA is a novel model for generating long-form audio from videos, leveraging the Diffusion Transformer (DiT) architecture. This project aims to address the challenges of generating long-form audio, which is crucial for applications in video editing and post-processing, and outperforms existing methods in both performance and effectiveness. ### Key Features: - **Long-Form Video-to-Audio Generation**: LoVA is designed to generate long-form audio, addressing the limitations of existing methods that focus on short-form audio. - **Diffusion Transformer Architecture**: Utilizes the Diffusion Transformer (DiT) architecture for improved performance in long-form audio generation. - **Comprehensive Evaluation**: Extensive objective and subjective experiments demonstrate that LoVA achieves comparable performance on short-form audio benchmarks and significantly outperforms all other baselines on long-form video inputs. ### Tech Stack: - **Programming Languages**: Python - **Dependencies**: Stable-Audio-Tools (for environment setup and model inference) ### Why It Matters: LoVA represents a significant advancement in the field of video-to-audio generation, providing a more effective and versatile solution for long-form audio synthesis. This project not only contributes to the advancement of audio synthesis techniques but also has practical applications in various fields such as video editing, post-processing, and more. ### Repository: - **Repository**: Ceaglex/LoVA - **GitHub Stars**: 16 - **GitHub Description**: The code and weight for LoVA. LoVA is a novel model for Long-form Video-to-Audio generation. Based on the Diffusion Transformer (