Projects

Overview: Nano Banana Pro API, also known as `gemini-3-pro-image-preview`, is an OpenAI-compatible endpoint offering image generation and editing services at a cost of $0.03 per image, with a billing model of $0.03 per image for 4K resolution. It provides access to 300+ image, video, and language models without requiring a monthly subscription. Key features: - Nano Banana Pro API supports 4K image generation and editing. - It offers a per-image pricing model with a $0.03 per image cost for 4K resolution. - The API supports a base URL and authorization header for seamless integration. - It supports multiple image URLs as input, allowing for reference images. - The API is designed for asynchronous processing, with tasks reported in the response. Tech stack: Python Why it matters: Nano Banana Pro API simplifies the process of integrating OpenAI-compatible image generation and editing services, providing a cost-effective and scalable solution for developers without the need for a monthly subscription or complex billing models.

### Overview: BioRAG is a biomimetic cognitive memory system designed for AI agents, featuring a hybrid PPMI+SVD vector search, LTP/LTD, and inter-agent shared cortex. It is implemented entirely in Python using SQLite, with no heavy dependencies or GPU usage. The system is modular, featuring 18 core modules for memory and 15 for server management, and supports 100-dimensional PPMI-SVD vectors and a semantic graph with 13 axes. ### Key Features: - **Hybrid Vector Search**: Combines PPMI and SVD for efficient and accurate retrieval. - **Biomimetic Design**: Inspired by biological systems, not a literal replica of a human brain. - **Semantics and Graphs**: Semantic understanding and a Hebbian synaptic graph for context preservation. - **Neuron Learning**: Self-learning neuron that reorders memories. - **Associative Neuron**: Closed factory neuron for associative learning. ### Tech Stack: - Pure Python + SQLite, zero heavy dependencies. ### Why It Matters: BioRAG provides a robust and context-preserving memory system for AI agents, enhancing their ability to maintain context across sessions. The modular architecture and semantic understanding contribute to its effectiveness in various AI applications.

Overview: This repository documents my learning journey in Deep Learning, featuring notebooks and projects on various architectures and frameworks. Key features: - Custom-built model architectures - Jupyter notebooks covering deep learning topics - Helper functions for data loading, visualization, and metrics - Technologies used: Python 3, TensorFlow/Keras, PyTorch, NumPy, Pandas, Matplotlib, Seaborn Tech stack: Python, TensorFlow/Keras, PyTorch, NumPy, Pandas, Matplotlib, Seaborn Why it matters: This repository helps developers understand and apply Deep Learning concepts, from fundamentals to advanced topics like CNNs, RNNs, Transformers, and more. Repository: diaazg/Deep-learning GitHub stars: 9 GitHub description: Deep learning notebooks and model implementations, exploring CNNs, RNNs, LSTMs, Transformers, and more during my AI learning journey.

### Overview: GPT3DataGen is a Python package designed to generate fake data for fine-tuning OpenAI models. ### Key Features: - **Data Generation**: Generates fake data for various model types, including classification, completion, and more. - **Flexible Formats**: Supports JSONL, JSON, CSV, and TSV formats for easy integration. - **CLI Interface**: Provides command-line options for easy configuration and execution. - **Install & Usage Guide**: Detailed instructions on how to install and use the package. ### Tech Stack: Python ### Why it Matters: GPT3DataGen simplifies the process of generating high-quality fake data for fine-tuning models, making it easier to experiment with and improve model performance. ### Repository: donwany/gpt3datagen ### Primary Language: Python ### GitHub Stars: 9 ### GitHub Description: GPT3DataGen is a python package that generates fake data for fine-tuning openai models.

### Overview: This project provides an official implementation of the paper "WMARK@ICLR: Detection Limits and Statistical Separability of Tree Ring Watermarks in Rectified Flow-based Text-to-Image Generation Models." It focuses on detecting and verifying tree ring watermarks in rectified flow-based text-to-image generation models. ### Key features: - **Detection Limits and Statistical Separability:** The project includes methods to determine the detection limits and statistical separability of tree ring watermarks in rectified flow-based models. - **Rectified Flow-based Text-to-Image Generation Models:** It utilizes rectified flow-based models for text-to-image generation, ensuring the integrity and authenticity of the generated images. - **Official Implementation:** The project provides an official implementation of the paper, making it accessible for further analysis and verification. ### Tech stack: Jupyter Notebook ### Why it matters: This project is crucial for researchers and developers interested in the detection and verification of tree ring watermarks in rectified flow-based models, ensuring the authenticity and integrity of generated images in applications such as text-to-image generation and watermark detection.

Overview: This project presents a fine-tuned implementation of the XLM-RoBERTa base model, a multilingual transformer-based language model trained on 100 languages from Hugging Face, specifically tuned for Vietnamese language question answering. Key features: - Fine-tuned XLM-RoBERTa model for Vietnamese question answering - Utilizes a self-attention mechanism for capturing long-range dependencies - Leverages a combination of translated datasets and a Vietnamese text dataset for training Tech stack: Jupyter Notebook Why it matters: This project showcases the potential of transformer-based models and multilingual models in natural language processing tasks, particularly in Vietnamese language question answering. Repository: ducnh279/Context-based-Question-Answering-for-Vietnamese Primary language: Jupyter Notebook GitHub stars: 9 GitHub description: Context-based-Question-Answering-for-Vietnamese README excerpt: # 🌟 Introduction to Vietnamese Question Answering This project presents a fine-tuned implementation of the **XLM-RoBERTa** base model, which is a multilingual transformer-based language model trained on 100 languages from Hugging Face. The model is specifically fine-tuned for Vietnamese language question answering task. # 💡 Model The XLM-RoBERTa base model, a powerful multilingual transformer-based language model that has demonstrated exceptional performance in various natural language processing tasks, is utilized in the this project. It leverages a self-attention mechanism to capture long-range dependencies and acquire knowledge of complex relationships between words and sentences. The

### Overview Modern Mixture-of-Experts (MoE) language models can have over 100 billion parameters, but they often exceed the available RAM on low-RAM machines. Swap-MoE is a patch set for the llama.cpp library that allows running MoE models on a machine with 16GB of RAM without needing a GPU. By using `mmap` to load only the active parameters, Swap-MoE enables inference on MoE models that would otherwise fail to load due to exceeding the RAM limit. ### Key Features - **Expert Streaming**: Only the active parameters of MoE models are loaded into RAM using `mmap`, allowing inference on models that exceed the RAM limit. - **LRU Keep-Recent**: Controls the retention of recently used experts in RAM. - **Expert Fixation**: Prefetches experts that are likely to be selected by the router. - **Expert Prefetch**: Prefetches experts that are likely to be selected by the router. ### Tech Stack - **C++**: The primary language used for development. - **LLama.cpp**: The llama.cpp library that is patched to enable MoE inference. ### Why It Matters Swap-MoE enables inference on MoE models that would otherwise fail to load due to exceeding the RAM limit, making it possible to run large MoE models on low-RAM machines without the need for a GPU. This is particularly useful for developers and researchers who need to run MoE models on constrained hardware. ### Repository - **Repository**: ek1507280

### Overview: ElBruno.Text2Image is a .NET library designed for local image captioning using ONNX Runtime. It automatically downloads HuggingFace models and supports various providers like FLUX, GPT-Image, and MAI-Image-2.5. ### Key Features: - **Automatic Model Download**: ElBruno.Text2Image downloads HuggingFace models automatically, making it easy to use without manual intervention. - **Image Captioning**: The library generates captions for images, enhancing the visual content. - **Providers Support**: Supports providers like FLUX, GPT-Image, and MAI-Image-2.5, offering flexibility in image processing. ### Tech Stack: - **Language**: C# - **GitHub Stars**: 9 ### Why It Matters: ElBruno.Text2Image simplifies the process of image captioning by automating the model download and providing a straightforward interface. This makes it particularly useful for developers looking to integrate image captioning capabilities into their applications without the hassle of manual model management. ### Repository: - **Repository**: elbruno/ElBruno.Text2Image - **GitHub Description**: A .NET library for local image captioning using ONNX Runtime with automatic HuggingFace model download - **GitHub Stars**: 9

Overview: Servekit is a toolkit designed to optimize the cold start process for large language models, specifically targeting LLMs with minimal latency. It wraps the SGLang library and implements optimizations to speed up weight loading and overall cold start times. Key features: - Fast weight loading for Lustre storage systems like CSCS’s `capstor` and `iopsstor`. - Pipeline parallelism for improved performance. - Support for multi-node fast weight loading. - Pipeline parallelism for improved performance. Tech stack: Python. Why it matters: Servekit significantly reduces cold start latency, making it an essential tool for developers aiming to deploy LLMs quickly and efficiently.

Overview: gpt-image-cookbook is a comprehensive multi-provider AI image generation toolkit that includes a curated prompt gallery, an agentic skill, and a Python CLI. Key features: - A curated prompt gallery - An agentic skill for Claude Code, Codex, OpenClaw, and Hermes - A CLI for switching between OpenAI gpt-image-2, Google Imagen, and Flux Tech stack: Python Why it matters: This toolkit provides pre-made prompts that work across multiple AI image generation models, simplifying the process of generating high-quality images and maintaining consistency across different providers. Repository: eugeniughelbur/gpt-image-cookbook GitHub stars: 9 GitHub description: Multi-provider AI image generation cookbook: prompt gallery, agentic skill, and CLI for OpenAI gpt-image-2, Google Imagen, Flux, and more. Claude Code plugin + Codex/OpenClaw skill + Python CLI.

### Overview: Discontext is a Discord bot designed to help users find messages semantically within their server. It uses OpenAI's Ada model to generate embeddings and stores them in a Qdrant vector database for efficient search. ### Key features: - Semantic search for Discord messages - Vector database for storing and searching messages - Supports filtering by author, channel, and date - Embeds for semantic similarity - No support for messages created before the bot was added to the server ### Tech stack: TypeScript ### Why it matters: Discontext aims to provide a useful tool for users who want to quickly find specific messages in their Discord server, even if they can't recall the exact phrasing. By leveraging vector databases and embedding models, it aims to offer a more efficient and accurate search experience compared to Discord's built-in search functionality. Repository: f0lio/discontext Primary language: TypeScript GitHub stars: 9 GitHub description: Semantic search for Discord messages.

### Overview: Lumen is a Rust-based inference engine for large language models, designed specifically for Apple Silicon and NVIDIA CUDA. It provides a single binary that downloads, converts, and runs models on the backend of your choice, ensuring zero ML dependencies and native performance. ### Key Features: - **One Self-Contained Binary**: Lumen is a single binary that downloads, converts, and runs models on the backend of your choice. - **Zero ML Dependencies**: No PyTorch, ONNX, or Python dependencies. - **GPU-Resident Execution**: Models are converted and run on the backend's GPU, ensuring fast and efficient decoding. - **Native Kernels**: Native CUDA C and Metal MSL kernels, a native BPE tokenizer, and the `.lbc` model format. - **OpenAI- and Anthropic-Compatible HTTP Server**: A zero-ML server that supports SSE streaming and template-driven tool calls. - **Text to Image**: A separate image-only server for text-to-image generation using Qwen-Image-2.1. - **Cross-Platform**: Runs on Apple Silicon (M-series) and NVIDIA CUDA (compute capability 7.5+). - **Tuned for Interactivity**: Single-stream, GPU-resident decode latency, and optimized for interactive serving. ### Tech Stack: - Rust ### Why it Matters: Lumen simplifies the deployment of large language models by providing a zero-ML dependency solution, ensuring fast and efficient inference on the backend of your choice. It is optimized for interactive serving, making it ideal for applications requiring real-time

### Overview: Feather is a lightweight, in-process memory layer designed specifically for agents, providing a compact and efficient way to store and retrieve information. Unlike traditional databases, Feather operates entirely within an agent's memory, eliminating the need for a separate server or network hop. This makes it ideal for real-time applications where performance and simplicity are crucial. ### Key Features: - **In-process Memory**: Feathers store is integrated within an agent’s memory, eliminating the need for a separate server or network hop. - **Adaptive Decay**: Memories age slower as they are recalled, ensuring that the most recently used data is retained. - **Trust Levels and Confidence**: Separate from importance, trust levels and confidence scores can be filtered and managed. - **Scoped Working Sets**: Memories are scoped to specific contexts, with hot sets and dotted scopes to optimize memory usage. - **Context Packets**: Required constraints are resolved first, and the memory fails closed if one is missing. - **Expiry**: Per-record `ttl` ensures that short-lived agent runs clean up their memory. ### Tech Stack: Feather is built using Python, leveraging the Python ecosystem for its development and deployment. ### Why It Matters: Feather is designed to be a compact and efficient solution for agents, providing a way to store and retrieve information that is both in-process and scalable. Its lightweight nature makes it ideal for real-time applications where performance and simplicity are key.

Overview: FrogDB is a fully Redis 8.x compatible database built in Rust, designed for speed, safety, and ease of operation. It supports replication, clustering, persistence, and additional features like vector sets, geospatial indexing, and pub/sub. Key features: - Redis 8.x compatibility - Cross-slot operations - Raft-based consensus for cluster coordination - Persistence using RocksDB - Load testing and benchmarking - Extensive testing suite - Self-hosted runner for running GitHub Actions locally Tech stack: Rust, Tokio Why it matters: FrogDB offers a faster, safer, and more scalable alternative to existing Redis solutions, with a solid foundation built on Rust and Tokio, making it easier to operate and maintain. Repository: frogdb/frogdb GitHub stars: 9 GitHub description: Fully Redis 8.x compatible database complete with replication, clustering, persistance, and hot/warm tiering.

### Overview: This repository contains a curated list of the best AI applications and tools, organized into 11 categories. The repository features 127 verified picks, updated regularly, and kept lean with only apps that genuinely stand out. ### Key features: - AI Chatbots - AI Coding - AI Search & Research - AI Search Visibility - AI Writing - AI Image - AI Video - AI Audio & Voice - AI Productivity & Knowledge - AI Automation & Agents - Local & Open Source ### Tech stack: Python ### Why it matters: This curated list is essential for developers looking to enhance their productivity and creativity with AI-powered tools. The repository provides a comprehensive overview of the best AI applications, ensuring that users can quickly find and integrate the most suitable tools for their projects.

Overview: Insight-Chat is a document chat application that allows users to have conversations with their documents using LlamaIndex and MongoDB Atlas. Key features: - Chat with documents - Ingest data from files, websites, or YouTube - Manage messages, chat sessions, and user information - Use LlamaIndex’s storage system for efficient document management Tech stack: Python, FastAPI, LlamaIndex, MongoDB Atlas, Atlas Vector Search Why it matters: Insight-Chat leverages cutting-edge technology to provide a powerful and efficient way to interact with documents, making it a valuable tool for developers and content creators alike.

Overview: Docsally is an experimental PDF assistant powered by Ollama, designed to instantly understand and answer questions based on your documents. Key features: - Instant document understanding and question answering - Vector database storage for smart vector queries - Easy installation and use, suitable for beginners Tech stack: JavaScript Why it matters: Docsally simplifies the process of understanding and summarizing documents, making it accessible and effective for users of all skill levels. Repository: hish22/docsally Primary language: JavaScript GitHub stars: 9

### Overview: IELTS Writing Assistant using LLMs and special prompting techniques. ### Key features: - Uses LLMs for text generation and chart/diagram understanding. - Two agents for Writing Task 1 and 2. - Self-understanding and self-correction for feedback. - Operates on Google Colab. ### Tech stack: Jupyter Notebook ### Why it matters: This project leverages the power of large language models to grade IELTS essays, providing constructive feedback and improving the evaluation process.

### Overview: This is a Linux kernel patch series designed to connect two AMD Strix Halo Linux machines with low-overhead USB4 data exchange for distributed inference. It provides USB4 stream character devices, an experimental DMA-BUF GPU-direct path, and fixes for AMD USB4 controller lifecycle/recovery. The series is experimental, out-of-tree work, and not submitted upstream. It applies to Linux version 7.3-rc5 and is tested on one physical pair of Ryzen AI Max+ 395 machines. ### Key Features: - **USB4STREAM Character Devices:** Batched submission, NAPI-style RX completion, and per-stream latency controls. - **DMA-BUF GPU Stream Extension:** Safe recovery, bounded queues, `sync_file` fences, and fail-closed recovery. - **Lifecycle, Teardown, and Recovery Fixes:** Fixes for the AMD USB4 NHI found while running the above on real hardware. ### Tech Stack: C ### Why It Matters: This series is a research kernel patch designed for a specific hardware topology and Linux kernel version. It provides a low-overhead data exchange mechanism for distributed inference tasks, which is not a generic USB4 replacement driver or a plug-and-play networking product.

### Overview: DeepSeek-V4.1-Flash is a high-performance model serving system designed for eight NVIDIA DGX Spark nodes. It utilizes tensor parallelism and DSpark speculative decoding, achieving 1 million context serving on a 200GB RoCE fabric. This system is a four-node recipe scaled to eight nodes, with the companion job to the original glm-5.3-uncensored-8x-dgx-spark. The model is optimized for vision and tool calling, with measurements and a detailed boot log available. ### Key features: - **Tensor Parallelism**: Utilizes 8 nodes for tensor parallelism. - **DSpark Speculative Decoding**: Enhances decoding efficiency. - **CUDA Graphs**: Optimized for faster inference. - **Vision and Tool Calling**: Integrated for improved performance. - **Context Serving**: Achieves 1 million context serving on 8x DGX Spark nodes. ### Tech stack: Shell-based system optimized for performance and efficiency. ### Why it matters: This system is designed for high-performance computing environments, leveraging advanced GPU technology to achieve significant improvements in processing speed and resource utilization. It is particularly useful for applications requiring large-scale data processing and real-time analysis, such as in AI research and development. ### GitHub stars: 9 stars ### Repository: im0xMagnus/deepseek-v4.1-flash-uncensored-8x-dgx-spark