Projects

Overview: Nano Banana 2 Lite API, developed by OpenAI-compatible APIMart, offers a cost-effective alternative to Gemini 3.1 Flash Lite Image, providing access to 300+ models and a single endpoint for $0.0125 per image. Key features: - Supports 300+ OpenAI-compatible models - Single endpoint for $0.0125 per image - Compatible with API-Key authentication - Pay-as-you-go pricing model - Asynchronous execution with task ID tracking Tech stack: Python Why it matters: Nano Banana 2 Lite API simplifies access to a vast array of models at a more affordable rate, making it ideal for developers and businesses looking to integrate AI into their projects without the high costs associated with traditional APIs. Repository: k3idrzqs5wbuf/nano-banana-2-lite-api-de Primary language: Python GitHub stars: 16 GitHub description: Bananen-KI / Nano-Banane — Nano Banana 2 Lite API (`gemini-3.1-flash-lite-image-ext`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: Bananen-KI / Nano-Banane — Nano Banana 2 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-3.1-flash-image-preview`) and the cost per image. ### Key features: - 8 tested recipes - Example renders, aspect ratios, model id, and cost per image - Nano Banana 2 (`gemini-3.1-flash-image-preview`) with 300+ other models - $0.015 per image (0.5K, 1K, 2K) - OpenAI-compatible endpoint - No monthly subscription, pay-as-you-go - Real-time billing - Asynchronous design ### Tech stack: Python ### Why it matters: - Key for Nano Banana 2 (`gemini-3.1-flash-image-preview`) usage with APIMart - $0.015 per image (0.5K, 1K, 2K) - OpenAI-compatible endpoint - No monthly subscription, pay-as-you-go - Real-time billing - Asynchronous design

### Overview: Seedance 2.0 Fast Prompts is a comprehensive guide to creating high-quality image prompts for the Seedance 2.0 Fast model, offering 8 tested recipes with detailed examples, aspect ratios, and pricing information. ### Key Features: - **4 Recipes**: Detailed prompts for various subjects, lighting, framing, and settings. - **Example Renders**: Visual examples of each recipe, including product ads, cinematic scenes, and food commercials. - **Pricing**: Estimated cost per second on the APIMart route, with a monthly subscription option. ### Tech Stack: - Python ### Why It Matters: - **Efficient Usage**: Uses a single API key to access multiple models, eliminating the need for pre-paid plans. - **Cost-Effective**: Charges only for the actual usage, making it a cost-effective solution. - **Detailed Prompt Structure**: Provides a cheat-sheet for creating effective prompts, including specific elements like subject, light, framing, and setting. ### Repository: Repository: k9fc4/seedance-2-0-fast-prompts-ko GitHub Stars: 16 GitHub Description: Seedance 2.0 Fast Prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`doubao-seedance-2.0-fast`) and the cost per image.

Overview: I-HallA-v1.0 is an official implementation of the paper "Evaluating Image Hallucination in Text-to-Image Generation with Question-Answering" presented at AAAI 2025. The project aims to evaluate image hallucination in text-to-image generation by measuring the factuality of generated images through visual question answering (VQA). Key features: - I-HallA (Image Hallucination evaluation with Question Answering) - I-HallA v1.0, a curated benchmark dataset - A three-stage pipeline for generating question-answer pairs using multiple GPT-4 Omni-based agents with human judgments Tech stack: Python Why it matters: I-HallA v1.0 provides a comprehensive evaluation metric for the accuracy of generated images in text-to-image generation, helping researchers and practitioners identify and address the issue of image hallucination.

Overview: SmartDoc is an open-source RAG application built with Gemini, Pinecone, and Next.js, designed for secure and powerful RAG systems with full control over document storage. Key features: - Document upload and query - RAG-powered question answering - Full data control - Secure authentication - Dark/Light mode toggle - Dynamic UI with animations - YouTube integration Tech stack: TypeScript, Next.js, Gemini, Pinecone, OAuth, TailwindCSS, Framer Motion, Auth.js Why it matters: SmartDoc offers a free, secure, and powerful RAG system with full control over document storage, making it ideal for anyone looking for a robust and customizable solution. Repository: kanugurajesh/SmartDoc GitHub stars: 16 GitHub description: A rag application to chat with your documents with the help of gemini, pinecone and Next.js

Overview: 即梦图片 — Seedream 5.0 Lite API (`seedream-5-0-lite`): 一个 OpenAI 兼容端点，提供模型 id、定价、参数和示例，适用于图像、视频、语言等 300+ 模型。 Key features: - 模型 id - 按量付费 - 费用随响应一并返回 - 原生异步设计 Tech stack: Python Why it matters: 一个 Key，整个目录。同一个 `https://api.apimart.ai/v1` base URL 和 `Authorization` 请求头即可访问 Seedream 5.0 Lite（`seedream-5-0-lite`）以及图像、视频、语言等 300+ 模型——只需切换 `model` 字段，客户端无需改动。

### Overview: Qwen-Image 3.0 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`qwen-image-3.0-pro`) and the cost per image. ### Key features: - **Model ID:** `qwen-image-3.0-pro` - **Aspect Ratios:** 1:1, 3:4, 4:3, 8:5, 8:7, 8:10, 8:13, 8:16 - **Cost per Image:** $0.0286 ### Tech stack: Python ### Why it matters: Qwen-Image 3.0 Pro is a powerful tool for generating high-quality images with a focus on realism and detail. The prompts provided here offer a comprehensive set of recipes that can help developers create professional-grade images quickly and efficiently.

### Overview: Grok Imagine 1.5 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`grok-imagine-1.5-apimart`) and the cost per image. ### Key features: - **Model ID:** `grok-imagine-1.5-apimart` - **Aspect Ratios:** 1:1, 16:9, 4:3, 3:4, 85mm, 24mm, Macro, Overhead Flat Lay - **Cost per Image:** $0.015 ### Tech Stack: OpenAI-compatible endpoint `https://api.apimart.ai/v1` ### Why it Matters: - **Key:** Same `https://api.apimart.ai/v1` base URL and `Authorization` header for Grok Imagine 1.5 (via `grok-imagine-1.5-apimart`) and 300+ models. - **Cost:** $1 or less per API call, with real-time cost reported on each task. - **Async:** Batch processing and retries are handled by a simple queue system, no need for custom integrations. - **Price:** Estimated cost of the set: $0.12 for 8 recipes. ### Overview: This repository contains 8 tested Grok Imagine 1.5 image prompt recipes, each with an example render, aspect ratio, and the cost on the current route. The recipes include: - **Model:** `grok-imagine-1

Overview: This course is designed to guide you through the LLMOps pipeline, covering data management, automation, and deployment. It includes pre-processing training data for supervised fine-tuning and deploying custom Large Language Models (LLMs). Key features: - Data management - Automation and deployment - Pre-processing training data - Versioning data and models - Configuring and executing a supervised tuning pipeline - Monitoring model behavior Tech stack: Jupyter Notebook Why it matters: This course provides essential knowledge and practical skills for adapting an open-source pipeline to fine-tune LLMs for specific applications, ensuring responsible AI practices and effective data management. Repository: ksm26/LLMOps Primary language: Jupyter Notebook GitHub stars: 16 GitHub description: In this course navigates through the LLMOps pipeline, enabling you to preprocess training data for supervised fine-tuning and deploy custom Large Language Models (LLMs).

Overview: TIA Portal Translator is a modern, async-powered tool designed to translate TIA Portal texts exported to Excel using free Google Cloud Translate, OpenAI/Openrouter API, and DeepL APIs. This script includes legacy scripts and a wrapper for easier integration. Key features: - Translation using Google Cloud Translate, OpenAI/Openrouter API, and DeepL APIs - Supports multiple languages and text formats - Preserves legacy scripts and a wrapper for easy integration - Supports both synchronous and asynchronous execution Tech stack: Python Why it matters: This tool simplifies the process of translating TIA Portal texts, making it accessible and efficient for various projects and applications.

Overview: This project is a Model Context Protocol (MCP) server designed to interact with Nautobot APIs, providing semantic search, dynamic API requests, and comprehensive knowledge base access. It supports multiple Nautobot environments and includes tools for managing the knowledge base. Key features: - Dynamic API Access - Semantic Endpoint Discovery - Knowledge Base Search - Multi-Environment Support - Smart Caching Tech stack: Python Why it matters: This MCP server simplifies interaction with Nautobot, making it easier to manage and access Nautobot-related data and knowledge. Repository: kvncampos/nautobot_mcp GitHub stars: 16 GitHub description: Nautobot Model Context Protocol (MCP) Server - Contains STDIO and HTTP Deployments with Embedding Search and RAG.

Overview: This project is a voice bot based on Large Language Models (LLM) that converts text to speech and vice versa, facilitating communication through voice. Key features: - Text-to-Speech (TTS) - Converts written text into spoken voice. - Speech-to-Text (STT) - Transcribes spoken text into written text. - Integration with AWS services: Uses Amazon Transcribe Streaming SDK for TTS, Amazon API Gateway for STT, and Amazon ElastiCache for Redis integration. Tech stack: Python Why it matters: This project demonstrates the integration of LLMs with voice-based communication, leveraging AWS services to streamline the process of converting text to speech and vice versa, making it accessible and efficient for developers and users alike. Repository: kyopark2014/llm-voicebot Primary language: Python GitHub stars: 16 GitHub description: It is a voice bot based on LLM. README excerpt: # 한국어 Voice-to-Voice 방식의 Voice Bot 만들기 ## Speech-to-Text (Voice Interpreter) Voice Interpreter extracts text from voice using the Amazon Transcribe Streaming SDK. The Python file `mic_main.py` is installed and executed on the device. If the device is an IoT device, it can be deployed as a Greengrass Component. This allows the device to extract text from the user's voice and utilize it in an Application based on the generated AI. The audio stream is extracted using `sounddevice` and the following region is specified to retrieve the text stream: `TranscribeStreamingClient

### Overview: The vLLM-Moet-GB10 project is a hardware adaptation of the 2-bit tensor-symmetric expert planes and FP4 delta tier, designed for a single NVIDIA GB10 system. This project aims to run large MoE models on a single GPU, achieving high throughput and high-quality results. ### Key features: - **2-bit expert planes**: The routed experts are ~97% of a large MoE's parameters, and each token uses only a few of them. They are stored as codes {-4, -1, 1, 4} times one scale per 32 weights, and hand-written SASS GEMMs compute directly on them. - **FP4 delta tier**: A pool of the hottest experts, served in FP4 instead of 2-bit and chosen at runtime from the actual routing. The full FP4 copy lives on NVMe, and promotions read from it while the model serves. - **Persistent planes cache**: Quantization runs once, on the first boot; every later boot loads the planes directly. - **Speculative decoding**: DSpark (DeepSeek's built-in head) for DS4, the DFlash2 block-diffusion drafter for GLM-5.3-Flash. - **Unified-memory engineering**: This fork includes streaming plane builds, the FP4 store on NVMe instead of pinned host RAM, page-cache handling, memory guards, and the runner hooks ported to vLLM's V2 model runner. ### Tech stack: vLLM ### Why it matters

### Overview: This project aims to develop a Multi-Agent RAG platform that utilizes Vector DBs and AI models to accurately answer AML/CFT compliance queries. The platform ensures quality validation and consistency checking through a robust tech stack. ### Key Features: - Multi-Agent Workflow - Streamlit App Interface - Integration with Qdrant for Vector DBs - OpenAI and LangChain for AI Model Utilization - FastAPI for API Server - Pydantic for Data Validation - Docker for Containerization - Streamlit for Streamlit App ### Tech Stack: The tech stack includes Python, OpenAI, LangChain, Qdrant, Docker, FastAPI, Streamlit, and Pydantic. ### Why It Matters: This project is crucial for financial institutions looking to enhance their AML/CFT compliance efforts by providing accurate and contextually relevant answers to compliance queries. The use of AI and Vector DBs ensures that the platform can handle complex, domain-specific data and provides a more reliable and consistent approach to compliance. Repository: luuisotorres/AML-MultiAgent-RAG GitHub Stars: 16 GitHub Description: A Multi-Agent RAG platform using Vector DBs and AI models to accurately answer AML/CFT compliance queries.

Overview: This project showcases the cutting-edge capabilities of Redis for vector similarity search, hybrid search, semantic caching, and advanced RAG models integrated with a language model (LLM) chatbot. It provides a comprehensive demonstration of how to leverage Redis as a vector database, enabling powerful applications across various domains. Key features: - Vector Similarity Search - Hybrid Search (Vector Similarity + Meta Search) - Semantic Caching - Advanced RAG Models for LLM Chatbots Tech stack: Python Why it matters: This project demonstrates the potential of Redis as a vector database, showcasing its versatility in various use cases such as vector similarity search, recommendation engines, semantic caching, and integrating with advanced models like RAG for LLM chatbots, thereby highlighting the importance of Redis in modern data processing and search technologies.

Overview: `pg_turboquant` is a PostgreSQL extension that adds a custom ANN index access method, `turboquant`, for compact nearest-neighbor search over `vector` and `halfvec`. Key features: - Custom access method: `USING turboquant` - Input types: `vector`, `halfvec` - Metrics: cosine, inner product, L2 - Modes: flat scan, IVF-routed scan, bitmap-filter support - Fast-path scope: faithful `v2` code-domain, compatibility fallback - SQL helpers: `tq_rerank_candidates`, `tq_approx_candidates`, `tq_recommended_query_knobs`, `tq_index_metadata`, `tq_maintain_index`, `tq_last_scan_stats` Tech stack: C Why it matters: `pg_turboquant` optimizes for compact storage and cache-friendly scoring within PostgreSQL, providing a denser index format for PostgreSQL users already storing embeddings with `pgvector`. GitHub stars: 16

Overview: A curated list of resources for Grok, the AI developed by xAI. Key features: - A comprehensive collection of tutorials and guides - Access to the latest updates and news - Expert interviews and case studies Tech stack: Python, JavaScript, HTML, CSS Why it matters: This repository serves as a valuable resource for developers interested in learning about Grok, the AI developed by xAI.

Overview: This project combines a fine-tuned LLaMa-3 model with LoRA adapters on 10-K Q&A data, an SEC data pipeline for real-time filings retrieval, and an end-to-end RAG pipeline for context-aware answers. It uses local embeddings and in-memory vector storage for efficient and accurate financial analysis. Key features: - Fine-tuned LLaMa-3 model with LoRA adapters - SEC API integration for real-time 10-K filings retrieval - Semantic search using open-source embeddings - End-to-end RAG pipeline implementation Tech stack: Python, Jupyter Notebook, Hugging Face Transformers, Sec API, LangChain, Sentence Transformers, Faiss-GPU-CU12 Why it matters: This system provides parameter-efficient fine-tuning, real-time data retrieval, semantic search, and context-aware answers, making it a powerful tool for financial analysts seeking accurate and relevant insights from SEC filings.

### Overview: This repository contains a comprehensive list of papers accepted at the 2024 European Conference on Computer Vision (ECCV). The papers cover various aspects of diffusion models, including motion generation, 3D point cloud denoising, and neural radiance field inpainting. The papers are categorized into different categories such as SMooDi, SV3D, EMDM, Diffusion Bridges, VFusion3D, Beta-Tuned Timestep Diffusion Model, Taming Latent Diffusion Model, FreeInit, LN3Diff, and UniDream. Each paper provides detailed information on the model architecture, results, and potential applications. ### Key Features: - **SMooDi**: A stylized motion diffusion model. - **SV3D**: A novel multi-view synthesis and 3D generation model. - **EMDM**: An efficient motion diffusion model for high-quality human motion generation. - **Diffusion Bridges**: A method for 3D point cloud denoising. - **VFusion3D**: A scalable 3D generative model from video diffusion models. - **Beta-Tuned Timestep Diffusion Model**: A model for neural radiance field inpainting. - **FreeInit**: A model to bridge the initialization gap in video diffusion models. - **LN3Diff**: A scalable latent neural fields diffusion model for 3D generation. - **UniDream**: A unified diffusion prior model for relightable text-to-3D generation. ### Tech Stack: The repository uses a combination of Python, JavaScript, and Hugging Face for

Overview: This is the official JavaScript SDK for the Modzy Machine Learning Operations (MLOps) Platform, designed to simplify interactions with the Modzy API. Key features: - Text-based model inference - File-based model inference - Embedded model inference - Database-based model inference - Cloud Storage-based model inference Tech stack: TypeScript Why it matters: This SDK streamlines the process of integrating Modzy's MLOps capabilities into JavaScript applications, making it easier for developers to leverage advanced AI and ML operations within their projects.