Projects

### Overview: This repository contains the official implementation of the paper **CCSR: Self-Rewarding Pretrained Text-to-Image Models**. The project focuses on improving text-to-image models by incorporating a self-rewarding mechanism that allows diffusion models to learn from their own generated images and continuously self-improve. ### Key Features: - **Self-Rewarding Mechanism**: This technique enables diffusion models to learn from their own generated images and continuously improve. - **Inspired by Self-Rewarding Language Models**: The method is inspired by the paper **Self-Rewarding Language Models**. - **Class-Conditional Self-Rewarding**: The mechanism is class-conditional, meaning it can be applied to specific classes or categories. - **LLM Model**: The self-rewarding mechanism is integrated with a large language model (LLM) for self-judging. ### Tech Stack: - **Jupyter Notebook**: The codebase is primarily written in Jupyter Notebook. ### Why It Matters: This project is significant for improving the quality and performance of text-to-image models by leveraging the self-rewarding mechanism, which is inspired by the advancements in self-rewarding language models. The ability to continuously learn and improve from generated images can lead to more accurate and diverse image generation in text-to-image models.

Overview: The AMD v620 Solutions Library is a containerized CI/CD build system designed to compile vLLM, llama.cpp, and HipFire from source against ROCm 7.x.x for AMD Radeon PRO V620 and RDNA2 GPUs. This library includes two separate images for base build, development, and validation tools, as well as specific images for inference and training workloads. Key features: - vLLM v0.28.x compilation - llama.cpp compilation - HIPFire compilation - ROCm 7.x.x base image - Development image with C tools dependencies - AMD GPU testing libraries and utilities - Compiled vLLM + PyTorch + Triton for RDNA2 inference - Compiled llama.cpp inference image for x86 with ROCm support - RDNA2 native inference engine with custom tuned kernels - Custom mixed-offloading inference engine for Qwen4 Arch - ModelScope Swift based training image for CPT, PEFT, LoRA, QLoRA workloads - PyTorch native training image with Transformers, Unsloth, PEFT, HuggingFace libraries Tech stack: Dockerfile Why it matters: This library streamlines the development and deployment process for AI inference and training on AMD GPUs, providing a consistent and efficient workflow for developers working with ROCm technology.

### Overview: AI-Resume is a containerized web application that transforms your CV into a conversational AI proxy. It provides a quick, semantic-based response to recruiters, allowing them to ask specific questions and receive detailed, evidence-based answers. ### Key Features: - **Semantic Search**: Uses natural language processing to provide accurate and contextually relevant answers. - **Real-Time Interaction**: Recruiters can interact with the AI agent in real-time, enhancing the interview process. - **Customizable Responses**: The AI can tailor its responses based on your semantic knowledge, making it highly personalized. ### Tech Stack: - **Python**: FastAPI for the backend API. - **Rust**: memvid gRPC for the semantic search and state management. - **React**: Frontend for the chat interface. ### Why It Matters: AI-Resume simplifies the process of screening candidates by providing recruiters with detailed, evidence-based answers. It streamlines the interview process, making it more efficient and effective. ### Repository: The repository is hosted on GitHub and has received 7 GitHub stars. The primary language used is Python, with TypeScript and Rust components for backend and frontend, respectively. ### Live Demo: Try the AI-Resume agent at [jane-doe-ai-resume.schwichtenberg.us](https://jane-doe-ai-resume.schwichtenberg.us/). This fictional candidate can be used to explore the agent's capabilities and functionality. ### Who Is This For: - **Resume Owner**: Deployed as an always-available AI agent, providing thoughtful answers about your experience without

### Overview: This project aims to develop and evaluate different models for binary classification tasks on the CBIS DDSM dataset, focusing on breast abnormalities classification and diagnosis. It utilizes various CNN architectures trained from scratch, pre-trained CNNs, a siamese network, and an ensemble technique. ### Key features: - **Binary Classification Tasks**: Breast abnormalities classification and diagnosis. - **Model Architectures**: Scratch CNN, Pretrained CNN (VGG, INC), Baseline CNN, Ensemble. - **Techniques**: CNNs, Siamese Networks, Ensemble Method. ### Tech stack: Jupyter Notebook ### Why it matters: This project provides a comprehensive framework for developing and evaluating different CNN models for binary classification tasks, showcasing the effectiveness of various techniques in improving classification accuracy and efficiency.

Overview: Shotit's core API, responsible for managing task coordination and image processing. Key features: - Image search request serving - Black borders cropping on search images - Rate limiting and user management - Index and database status serving - Compressed hash file serving - Hash job distribution to workers Tech stack: Node.js, MariaDB, Redis, Docker, Docker Compose, LIRESOLR, Milvus, MinIO, Etcd, Shotit Searcher, Shotit Sorter (optional), G++, CMake Why it matters: This API is essential for the Shotit platform, ensuring efficient task management and image processing through robust features and seamless integration with other Shotit components.

Overview: Imagify is an AI-powered image generation SaaS web application built with the MERN stack, utilizing the Clipdrop API for high-quality image creation. It includes a secure login/signup system, a user-friendly interface, and seamless integration between client and server. Key features: - AI-based image generation using the Clipdrop API. - User authentication system with login and signup functionality. - Clean and modern user interface. Tech stack: JavaScript, Node.js, Express.js, MongoDB, React, and Mongoose. Why it matters: Imagify offers a secure and efficient way to generate high-quality images through AI, making it valuable for businesses and individuals looking to create professional-looking visuals without the need for expensive software.

### Overview: Krill is a Swift-based local language model (LLM) engine and coding agent designed for macOS. It runs both a model and a coding agent within a single binary, providing a seamless experience without the need for Python. Krill is hosted on your Mac, meaning model traffic stays on the machine, and it supports multiple modes including chat, serving OpenAI APIs, and a full coding agent. ### Key Features: - **LLM Engine**: Runs a model directly on the local machine, leveraging Swift and MLX for efficient performance. - **Coding Agent**: Provides full-stack capabilities including text, vision, and voice functionalities. - **Multi-Mode Support**: Offers chat, serve, agent modes, and integrates with Claude Code and OpenCode for seamless integration. - **Web and Phone Access**: Supports web and phone access through the `/ui` endpoint, allowing sessions per repository and live tool calls. - **Tool Constrained**: The agent's tool names are constrained by the model's vocabulary, ensuring compatibility and preventing model failures. - **Continuous Batcher**: Optimized for throughput under load, with shared-prefix KV reuse and speculative decoding. - **Vision and Audio**: Utilizes native vision and audio capabilities, including SigLIP2 and USM Conformer. ### Tech Stack: Swift + MLX ### Why It Matters: Krill combines the best of both worlds, providing a powerful local LLM engine and a robust coding agent in a single binary. This approach eliminates the need for Python, making it ideal for developers who prefer Swift and macOS. The seamless integration with

### Overview: CoordiNode is a graph-native hybrid retrieval engine designed for AI applications and GraphRAG. It combines a graph database, a vector database, and a full-text search engine into a single, transactional engine with a single query language. ### Key Features: - **One Query**: Mixes graph traversal, vector similarity, and BM25 text relevance in a single pipeline. - **One Snapshot**: MVCC snapshot isolation across edges, embeddings, and the text index, ensuring a retrieval never reads partially applied writes. - **One Binary**: Runs as a single executable, eliminating the need for a JVM, C dependencies, or a separate router. ### Tech Stack: Rust - Built in Rust, zero garbage collection, single binary, openCypher-compatible, clustering included in the open-source edition. ### Why It Matters: - **Graph Database, Vector Store, and Search Engine**: All functionalities are integrated into a single engine, simplifying complex queries and reducing the need for multiple database setups. - **Single Query Language**: A unified query language simplifies development and reduces complexity. - **Zero GC**: Rust's zero garbage collection ensures efficient memory management. - **Single Executable**: The same binary runs standalone or as a Raft-replicated cluster, eliminating the need for a separate router or JVM. ### Repository: structured-world/coordinode - GitHub Stars: 7 - Primary Language: Rust - GitHub Description: The graph-native hybrid retrieval engine for AI and GraphRAG. - GitHub Releases: Latest release available via the GitHub repository. -

### Overview: This repository contains a collection of practical projects designed to deepen understanding of n8n's features and workflow automation capabilities. The projects are aimed at developers and include inventory management, order status tracking, RAG (Retrieval-Augmented Generation) with Pinecone and Supabase, customer support AI, and LinkedIn content creation workflows. ### Key Features: - **Inventory Agent**: Retrieves or updates inventory data from a Google Sheet. - **Order Status via Email**: Automatically sends email notifications based on order status from a Google Sheet. - **RAG - Uploading PDF to Pinecode Vector Database**: Implements a RAG pipeline for efficient retrieval from a Pinecone vector database. - **RAG - AI Agent**: Retrieves information from a Pinecone vector database to answer user questions. - **RAG with Supabase & Postgres**: Demonstrates RAG using Supabase and Postgres for efficient information retrieval. - **Customer Support AI Agent**: Automatically classifies emails and forwards them to an AI agent for response. - **LinkedIn Content Creator Agent**: Reads topics from a Google Sheet and generates concise summaries based on web searches. ### Tech Stack: n8n, Google Sheets, Pinecone, Supabase, OpenAI, Wikipedia, Gmail, Pinecone Vector Store, Supabase Vector Store, AI Agent, OpenAI Chat Model, Simple Memory, Recursive Character Text Splitter, Default Data Loader, Google Drive, Supabase, OpenAI Embeddings, Wikipedia, LinkedIn, Google Sheets, Pinecone, Supabase, Pinecone Vector Store, OpenAI Embed

### Overview LLAMA.CPP Manager is a modern GUI configuration tool designed for working with GGUF models. It provides a clean and intuitive interface to select and analyze models, configure inference parameters, and estimate VRAM usage. The application is lightweight, efficient, and cross-platform, making it ideal for developers and researchers. ### Key Features - **Ultra-lightweight**: <50 MB VRAM footprint - **Fast startup and single-binary deployment** - **Cross-platform**: Windows, macOS, Linux - **Clean Dear ImGui interface** - **Integrates with llama.cpp tools for accurate model analysis** - **Functional file browser for easy model selection** - **Export configurations to llama.cpp CLI commands** - **Optional native file dialog support** - **Recent models list with per-model settings persistence** - **Hybrid configuration storage for better organization** ### Tech Stack - **C++17 compatible compiler**: GCC 8+, Clang 7+, MSVC 2019+ - **CMake 3.15 or higher** - **SDL2** - **OpenGL 3.0+** ### Why It Matters - **Efficiency**: Streamlines the configuration process, reducing the need for manual adjustments. - **Accuracy**: Integrates with llama.cpp tools for precise model analysis. - **Cross-Platform**: Works seamlessly across different operating systems, ensuring compatibility and ease of use. - **Intuitive Interface**: Simplifies the selection and management of GGUF models, enhancing productivity. ### Installation - **Prerequisites**: Ensure you have a compatible C

Overview: TemRust-SMOL-v5-1.5B is a fine-tuned model of Qwen/Qwen2.5-Coder-1.5B-Instruct specialized for Rust code, trained on real GitHub PR fixes. It achieves a 67.6% pass rate on 37-task benchmark, surpassing the untrained base by 16.2 percentage points. Key features: - Fine-tuned model for Rust code fixes - 1.5B parameter count - 67.6% pass rate on 37-task benchmark - 83.8% pass rate in ensemble configuration - 70% pass rate for borrow-checker fixes - 77.8% pass rate for fix-this-bug tasks - 44.4% pass rate for test generation Tech stack: Python Why it matters: TemRust-SMOL-v5-1.5B provides a high-quality Rust code assistant, improving developer productivity and code quality through precise and effective fixes. Repository: temm1e-labs/temrust GitHub stars: 7

Overview: TinyInference is an incredibly lightweight inference module for TinyHumans AI agents, focused on model and embedding API concerns without owning an agent loop, graph runtime, middleware system, capability registry, or workspace policy. Key features: - Model and embedding API concerns - Chat model abstraction with real asynchronous streaming - OpenAI Chat Completions, OpenAI Responses, and OpenAI-compatible provider adapters for Anthropic, Ollama, DeepSeek, Groq, xAI, OpenRouter, Together, and Mistral - OpenAI, Cohere, Ollama, Voyage, cloud, no-op, and deterministic mock embeddings - Document reranking through Voyage, with original-position results, cancellation, bounded retries, and provider-reported usage - Request caching, stream accumulation, normalized provider failures, provider-neutral retry classification, and `Retry-After` parsing - Conservative context-window and vision-capability hints for raw model ids when a provider cannot supply an authoritative model profile - Normalized provider model-catalog parsing and local runtime model, vision, embedding, speech model, and voice resolution - Endpoint-only local inference against a runtime the user installs and runs (Ollama, LM Studio, MLX, OMLX, or any OpenAI-compatible server) - Reusable provider OAuth/PKCE, OpenAI Codex authentication, credential-file parsing, and deterministic provider-error classification - OpenAI-compatible hosted transcription, Piper synthesis, local-LLM transcript cleanup, and bounded PCM streaming helpers - `tinyinference-decisions`, typed Jev/System One and

Overview: Bananen-KI's Nano-Banane API provides access to a comprehensive set of OpenAI-compatible models, including 300+ image, video, and speech models, all under a single endpoint. This API offers a cost-effective alternative to traditional AIs, with a flat rate of $0.0137 per image and no monthly subscription needed. Key features: - Access to 300+ models - Flat rate pricing of $0.0137 per image - OpenAI-compatible endpoint - Supports text-to-image and reference-image editing - Asynchronous processing Tech stack: Python Why it matters: This API is ideal for developers looking for an alternative to traditional AI solutions, offering a cost-effective and flexible way to integrate AI into their projects without the need for a monthly subscription. GitHub stars: 7 Repository: tn-gr0cr/nano-banana-2-1-api-de

Overview: This repository provides a public, machine-readable record of LLM inference prices across 100+ platforms, updated every 6 hours. Key features: - Public price history for LLM inference across 100+ platforms. - Updated every 6 hours. - Git is used as the time-series database. Tech stack: Python. Why it matters: This repo fixes the issue of only some providers publishing today's price, ensuring a consistent record of LLM inference costs.

Overview: `cuda-kernel-optimizer` is a project designed to optimize the performance of ChatGPT's GPU workload by leveraging evidence-driven CUDA, CUTLASS, Triton, and GPU workload optimization techniques. It provides a structured approach to identify bottlenecks, evaluate candidate solutions, and ensure the results are both reliable and verifiable. Key features: - GPU workload optimization - Evidence-driven approach - Integration with ChatGPT's environment - Support for various workload types - Detailed analysis of kernel and execution paths Tech stack: Python Why it matters: By leveraging evidence and ChatGPT's expertise, `cuda-kernel-optimizer` ensures that the optimization process is both effective and verifiable, providing a robust solution for improving ChatGPT's performance on GPU workloads. Repository: troycheng/cuda-kernel-optimizer GitHub stars: 7 GitHub description: Evidence-driven CUDA, CUTLASS, Triton and GPU workload optimization for ChatGPT · Using ChatGPT to optimize GPU workload performance.

Overview: tsquare is a text-based wireframe board system that is optimized for low-fidelity screens, featuring easy writing and fast rendering capabilities. Key features: - Supports multiple devices side by side - Allows notes to be added beside screens - Renders as SVG or PNG for easy integration with LLMs Tech stack: TypeScript Why it matters: tsquare is designed to be easily written by LLMs, making it ideal for use with large language models for creating wireframes. Repository: tsquare-js/tsquare GitHub stars: 7 GitHub description: Wireframe boards in plain text: easy for LLMs to write, fast to render as SVG or PNG.

### Overview: Seedream 4.5 prompts are tested image prompt recipes that showcase eight effective and efficient methods for generating high-quality images. The prompts are designed to be easily replicable, with clear instructions on the model ID (`seedream-4-5`), aspect ratios, and the cost per image. ### Key Features: - **Model ID:** `seedream-4-5` - **Prompt Recipes:** Eight recipes listed, each with an example render, ratio, and cost on this route. - **Effective Price:** $0.026 per image, estimated cost of the whole set: $0.208. ### Tech Stack: Python ### Why It Matters: - **Versatile Access:** Seedream 4.5 can be accessed through a single `https://api.apimart.ai/v1` base URL and `Authorization` header, allowing for easy integration with catalog management systems. - **Minimal Setup:** No need for prepayment plans or monthly subscriptions; users can start with a single payment and adjust as needed. - **Clear Pricing:** All costs are reflected in the response, and failed tasks are refunded. - **Efficient Design:** The prompt structure is designed to be simple and effective, ensuring that the final output is as close to the intended vision as possible. ### Other Names: Seedream ### Prompts Overview: - **Model:** `seedream-4-5` - **Recipes:** Eight recipes, each with a clear example render, ratio, and cost on this route. - **Effective Price

Overview: Nano Banana Pro is an OpenAI-compatible API endpoint that provides image generation and editing services, offering a unique and cost-effective alternative to traditional image generation methods. Key features: - Nano Banana Pro (`gemini-3-pro-image-preview`) supports 4K image generation and editing with multi-step planning. - The API endpoint is compatible with OpenAI, allowing for seamless integration with existing systems. - It offers a per-image pricing model, with a minimum cost of $0.03 per image for 4K resolution. - The API supports both default and 4K pricing tiers, with the latter being slightly more expensive. - Users can access multiple models via a single base URL and API key, eliminating the need for a separate subscription or plan. Tech stack: Python Why it matters: Nano Banana Pro provides a cost-effective and efficient solution for image generation and editing, making it ideal for developers who need to integrate image-based features into their applications without the need for a dedicated API key or plan. GitHub stars: 7

### Overview: 即梦视频 — Seedance 1.0 Pro Fast prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`doubao-seedance-1-0-pro-fast`) and the cost per image. ### Key features: - **4 Tested Recipes**: Each recipe includes a concrete object or person, material, light direction, framing ratio, and setting. - **Effective Price**: $0.0079 per second, estimated cost of the set: $0.0316. - **Cost Per Image**: The cost is returned as part of the task result, and failed tasks are automatically refunded. ### Tech stack: Python ### Why it matters: - **Key, Entire Directory Access**: Same `https://api.apimart.ai/v1` base URL and `Authorization` header to access Seedance 1.0 Pro Fast (doubao-seedance-1-0-pro-fast) and 300+ models. - **Cost-Per-Second Pricing**: $0.0079 per second, no subscription or pre-paid plans required. - **Immediate Cost Reporting**: Each task reports the actual cost, and failed tasks are automatically refunded. - **Asynchronous Design**: Submit tasks, get `task_id`, and poll `GET /v1/tasks/{id}` for batch processing and retries without custom integration. ### Repository: v87tcvt6l/seedance-1-0-pro-fast-prompts-cn ### Primary Language: Python ### GitHub Stars: 7

### Overview: A multi-modal embedded database engine in pure Rust, designed for edge AI applications. It offers key-value, vector search, blob storage, TTL, merge operators, CDC, and flash storage backend all within a single ACID-transactional file. Key features include 95% recall@1 at 1,000+ QPS on SIFT1M, 1.6–3x faster writes than LMDB, and 2–3x faster reads than RocksDB. It is optimized for single-threaded, 8 GB RAM, and single-process, single-writer environments. ### Key features: - **Vector Search**: IVF-PQ with integer ADC, SIMD distance, and ACID-transactional index updates. - **Blob Storage**: Content-addressed deduplication, chunked writes, causal lineage graph. - **TTL Tables**: Per-key expiration for sessions. - **Merge Operators**: Built-in support for efficient data aggregation and joins. - **CDC (Change Data Capture)**: Built-in support for tracking changes in data. - **Flash Storage Backend**: Native vector search, blob store, TTL, merge operators, CDC, and flash storage backend. ### Tech stack: Pure Rust, no_std. ### Why it matters: - **Embedded Rust Applications**: Suitable for edge AI applications that require KV + vector search + blob storage in a single ACID-transactional process. - **Single-threaded**: Designed for single-threaded environments, making it efficient and lightweight. - **No C++ Dependency**: Avoids the