Projects

### Overview: This project profiles a set of 8 tested FLUX 2 Max prompts, including example renders, aspect ratios, and the cost per image on the `flux-2-max` route. The prompts are designed to guide developers in crafting effective image prompts for their applications. ### Key Features: - **8 Prompts:** Each recipe includes a concrete object, light direction, framing ratio, and setting details. - **Example Renders:** Illustrative images demonstrating how each prompt translates into a rendered image. - **Costs:** The cost per image is provided for each recipe, ranging from $0.056 to $0.104, depending on the aspect ratio. ### Tech Stack: - Python ### Why It Matters: - **Key 1:** The ability to access a catalog of 300+ models using a single base URL and `Authorization` header. - **Key 2:** The option to start at $1 per image, with no need for a subscription or pre-paid plan. - **Key 3:** Transparency in pricing, with all costs reported in the response, ensuring cost transparency. - **Key 4:** Non-blocking design, allowing for easy integration and management of tasks without complex integration points. ### Team Motivation: - **Key Motivation:** Simplify the process of using APIMart's FLUX 2 Max for developers, making it more accessible and efficient. ### Repository: - **Repository:** `nbkj6/flux-2-max-prompts-ja` - **Stars:** 8 - **Description

### Overview: OAI Chat is a multi-modal chatbot powered by OpenAI transformer models, designed for developers and users alike. It supports various file types and provides features like image upload, text file upload, streaming chat, and more. ### Key Features: - **Image Upload (via gpt-4-vision)**: Allows users to upload images and generate text descriptions. - **Word File (DOCX) Upload**: Enables users to upload and interact with Word documents. - **PDF File Support (via image rendering & GPT-4V)**: Converts PDFs into images for text-based interactions. - **Plaintext File Upload**: Allows users to upload plain text files. - **Chat History Download**: Enables users to download chat history for reference. - **File Download**: Supports downloading files created by the model, such as an ICS calendar file. - **Streaming Chat**: Enables real-time chat sessions. - **Image Generation (via gpt-image-1)**: Generates images based on textual descriptions. - **MCP Server Support (both remote and local)**: Allows for remote and local MCP servers. - **Optional UnrestrictedPython Execution**: Enables users to execute Python code with no restrictions. ### Tech Stack: Python ### Why it Matters: OAI Chat is a powerful tool for developers and users looking to integrate chatbot capabilities into their projects, utilizing OpenAI's transformer models for enhanced text generation and image processing.

Overview: This project aims to fine-tune Image Transformers using a learnable memory mechanism, as presented in the paper "Fine-tuning Image Transformers using Learnable Memory" by Sander et al. at CVPR 2022. Key features: - Implementing the paper's method in PyTorch - Reproducing the results with limited computational resources Tech stack: PyTorch Why it matters: This project aims to enhance the capacity of pre-trained models in a non-destructive manner, preserving performance on previous tasks while adapting to new ones.

Overview: This repository presents a Semantic Communications Framework utilizing Transformer-based models for text data extraction and reconstruction. Key features: - Semantic Information Extraction - Semantic Information Reconstruction - Named Entity Recognition (NER) - Relation Classification Tech stack: Python Why it matters: This framework provides a comprehensive solution for semantic communication by integrating advanced NLP models to extract and reconstruct text data, demonstrating its utility in various applications requiring semantic understanding and reconstruction.

Overview: This project applies Fine-tuning BERT & BioBERT on the BC5CDR dataset for biomedical named entity recognition, focusing on diseases and chemicals. Key features: - Fine-tuning BERT & BioBERT models - BC5CDR dataset for biomedical text - Extracting diseases and chemicals from biomedical text Tech stack: Python, Jupyter Notebook, Hugging Face Transformers Why it matters: This project demonstrates the effectiveness of fine-tuning techniques on biomedical NER, improving precision and recall by 27% and 26% respectively, making it a valuable resource for healthcare automation and clinical decision support.

Overview: Nano Banana 2 Lite API provides an OpenAI-compatible endpoint for generating images, flash images, and text-to-image conversion with a per-image pricing model. It offers a base URL and a single API key for easy integration. Key features: - Per-image pricing - Supports OpenAI-compatible endpoint - Model id and route details - Pricing details and effective price - Request parameters for prompt, size, resolution, and more Tech stack: Python Why it matters: Nano Banana 2 Lite API streamlines image generation by providing a single API key, making it cost-effective and easy to integrate with existing workflows. Repository: nkdje/nano-banana-2-lite-api-ru GitHub stars: 8

### Overview: Seedream 5.0 Lite Prompts: 8 Recipes with Outputs Seedream 5.0 Lite Prompts offer tested image prompt recipes with example renders, aspect ratios, the model id (`seedream-5-0-lite`), and the cost per image. This repository provides an overview of eight recipes, each accompanied by an example render, the aspect ratio, and the cost on this route. The effective price used is $0.028 per image, with an estimated cost of $0.224 for the entire set. ### Key Features: - **Model ID:** `seedream-5-0-lite` - **Recipes:** 8 (machine-readable in `data/prompts.json`, table in `PROMPTS.md`) - **Effective Price:** $0.028 per image (estimated cost: $0.224) ### Tech Stack: - Python ### Why It Matters: - **Cost Efficiency:** Seedream 5.0 Lite (`seedream-5-0-lite`) offers a cost-effective solution with a price of $0.028 per image, making it an ideal choice for those on a budget. - **Versatile Models:** The model supports over 300 different models of image, video, and language, ensuring a wide range of creative possibilities. - **Asynchronous Nature:** The API is designed to be asynchronous, allowing users to send requests and retrieve task IDs without needing to integrate with a specific API endpoint. ### Primary Language: - Python ### GitHub Stars

Overview: Qwen Image 3.0 API provides a cost-effective and OpenAI-compatible endpoint for generating high-quality images, text, and videos. Key features: - **Cost-Effective Pricing**: $0.0206 per image on the APIMart route. - **OpenAI-Compatible Endpoint**: Supports 300+ models including image, video, and language. - **Model Id and Routes**: Available via `qwen-image-3.0` model id with tiered pricing. - **Real-Time Pricing**: Reports actual cost per task and automatically refunds failed tasks. - **Batch Processing**: Supports batch and retries through simple queueing. Tech stack: Python Why it matters: Qwen Image 3.0 offers a cost-effective solution for generating high-quality images, making it ideal for developers and content creators looking to streamline their workflows without the need for expensive subscriptions or pre-paid plans.

### Overview: Imagina 4.0 prompts: 8 tested recipes with outputs, aspect ratios, and the model ID (`imagen-4.0-apimart`) and cost per image. This repository provides a comprehensive guide to crafting effective image prompts using the APIMart API, ensuring precise control over the rendering process and cost efficiency. ### Key Features: - **Model ID:** `imagen-4.0-apimart` - **Aspect Ratios:** 1:1, 4:3, 8:5, etc. - **Cost per Image:** $0.04 - **Prompt Structure:** - **Subject:** Concrete object or person with material - **Light:** Direction + Quality - **Framing:** Ratio + Lens or Distance - **Setting:** Surface, Backdrop, Environment - **Finish:** Render or Film Character ### Tech Stack: Python ### Why It Matters: - **Cost Efficiency:** The API offers a cost-effective solution for image generation, allowing users to pay only for the images they need. - **Precision Control:** By following specific prompts, users can ensure that their images meet their desired specifications, leading to more accurate and visually appealing results. - **Versatility:** The repository includes 8 different recipes, covering various subjects and lighting conditions, making it a versatile tool for developers and designers. ### Repository: - **Primary Language:** Python - **GitHub Stars:** 8 - **GitHub Description:** Imagen 4.0 prompts: tested image prompt recipes

### Overview: This is a reproducible SGLang container source for the official `Qwen/Qwen3.8-Flash-Next-FP8` checkpoint on two NVIDIA RTX PRO 6000 Blackwell GPUs (SM120). The repository includes the exact candidate release and is designed for dual-SM120 startup, text, reasoning, tool-call, image-input, decode, cold-prefill, full GSM8K, and five-run supported-concurrency publication panel. The model passed all the necessary benchmarks and reached near-native-context and non-speculative panels. ### Key features: - Dual-SM120 GPU support - Native 262,144-token context - NextN MTP-3: three steps, top-k 1, four draft tokens - FlashInfer GDN prefill/decode and QSA sparse attention - PLE table offloaded to pinned host memory - Text, image input, separated reasoning, and tool calls - HiCache disabled - Dedicated, bounded HiCache volume deferred to v0.2 ### Tech stack: Shell ### Why it matters: This source release is an experimental version of SGLang Qwen3.8-Flash-Next SM120, designed for dual-SM120 GPUs and optimized for performance and accuracy. It passes all necessary benchmarks and reaches near-native-context and non-speculative panels, making it suitable for applications requiring high performance and reliability.

### Overview: AI图片处理，包括高清扩图、人物头部提取、头像卡通化、线稿提取、图片上色、风格转换等功能，适用于多种场景，提升图片质量与艺术效果。 ### Key features: - 高清扩图 - 人物头部提取 - 头像卡通化 - 线稿提取 - 图片上色 - 风格转换 ### Tech stack: Vue ### Why it matters: AI图片处理技术在提升图片质量、艺术效果以及满足个性化需求方面具有显著优势，广泛应用于社交媒体、个人创作、广告设计等领域。

Overview: This project is a developer-friendly tool designed to monitor and manage the costs of self-hosted vLLM inference servers, providing insights into effective cost-per-million tokens and real-time telemetry. Key features: - Real-time monitoring of vLLM server performance metrics - Effective cost-per-million tokens meter - Support for live telemetry and effective cost-per-million tokens meter - Companion artifact to a paper on beyond per-token pricing methodology Tech stack: Python Why it matters: This tool offers a practical solution for developers managing vLLM servers, providing insights into cost efficiency and performance metrics without the complexities of serverless pricing models. Repository: pChitral/vllm-cost-meter GitHub stars: 8 GitHub description: Concurrency-aware effective-cost meter for self-hosted vLLM inference — companion artifact to arXiv:2606.11690 (Beyond Per-Token Pricing).

Overview: This repository provides production-ready checklists and frameworks for deploying Large Language Models (LLMs), diffusion models, embedding pipelines, and multimodal AI systems. It covers deployment, infrastructure setup, observability, security, and governance, ensuring reliable and efficient AI systems. Key features: - Deployment checklists - Infrastructure best practices - Runtime diagnostics - Compliance and governance frameworks Tech stack: Shell Why it matters: This repository helps teams build reliable, observable, scalable, and cost-efficient AI systems, addressing the unique challenges of deploying AI models, including non-linear batching behavior, GPU memory fragmentation, and performance optimization. Repository: paralleliq/piqc-knowledge-base GitHub stars: 8

Overview: The Canada Labour Research Assistant (CLaRA) is a privacy-first LLM-powered research assistant designed to directly quote sources to mitigate hallucinations and construct context-grounded answers about Canadian labour laws, standards, and regulations. It operates entirely offline, ensuring the confidentiality of your conversations. Key features: - Retrieval-Augmented Generation (RAG) to infuse context in each query. - Chunking strategy to improve question answering. - Metadata leveraging to improve question answering and make information easily verifiable. - Reranking to prioritize relevant sources when detecting legal provisions. - Dynamic context window allocation to prevent document chunks from being truncated and manage memory efficiently. - Performance optimizations to reduce latency (database caching, tokenizer caching, response streaming). - Runs locally and offline, ensuring confidentiality. - Minimalist base dependencies for better portability and resilience. - Bring-Your-Own-Model option with Ollama and vLLM. - Bring-Your-Own-Inference-Provider and switchable between local and remote modes in the UI. - RAG-enabled conversation history with previous document chunks for deeper context. - On-the-Fly LoRA Adapters for fine-tuned models with vLLM. - 3 runtime modes: normal, evaluation, and profiling. - Evaluation mode measures the quality of responses. - Profiling mode tracks the impact of each function call and component in the architecture. Tech stack: Python. Why it matters: CLaRA ensures the confidentiality of conversations by operating entirely offline, making it ideal for scenarios where data privacy is a priority.

Overview: This repository provides a solution for the Data Fusion Contest task 1, focusing on improving the accuracy of item category prediction in receipts using a distilbert model. Key features: - Pretrained on masked language modeling - Label smoothing - Custom model architecture (weighted sum of hidden states + multisample dropout) - Lowercase input - Increased number of training epochs - Pseudolabeling technique - Distilbert model finetuning Tech stack: Python, Jupyter Notebook Why it matters: The repository showcases advanced techniques for enhancing model performance in a real-world application, demonstrating the effectiveness of pseudolabeling and other strategies in improving prediction accuracy.

Overview: Refind is a project designed to store and search personal documents using Weaviate's vector database and OpenAI's vectorizer module. Key features: - Integrates Weaviate for vector storage - Utilizes OpenAI's vectorizer for efficient querying - Supports generative search capabilities Tech stack: Go Why it matters: This project showcases the integration of cutting-edge vector databases and AI technologies, providing a comprehensive solution for document management and search. Repository: r0mdau/refind Primary language: Go GitHub stars: 8 GitHub description: Experimenting Weaviate vector database with OpenAI vectorizer module and generative search

### Overview: Smart Fork Detection is a powerful tool designed to overcome the limitations of AI-assisted development by enabling semantic search and intelligent session forking, allowing users to seamlessly resume their work from the most relevant past conversations. ### Key Features: - **Semantic Search**: AI-powered search across all Claude Code sessions. - **Smart Session Forking**: Resume from the most relevant conversation. - **Background Indexing**: Automatic real-time indexing of new sessions. - **Project-Scoped Search**: Filter results by project directory. - **Fork History Tracking**: Keep track of recently forked sessions. ### Tech Stack: Python ### Why It Matters: Smart Fork Detection transforms AI-assisted development by enabling seamless context recovery and knowledge reuse, significantly boosting productivity and efficiency. ### Repository: recursive-vibe/smart-fork ### GitHub Stars: 8 ### Primary Language: Python ### GitHub Description: Smart Fork Detection - MCP server for Claude Code that enables semantic search of past session transcripts and intelligent session forking. Find and resume relevant conversations instantly with AI-powered search, clustering, and session management. ### Table of Contents: - Installation - Quick Start - Usage - Using the fork-detect Tool - Selecting a Session - Forking a Session - Configuration - Configuration Options - Batch Mode Setup - Configuration File - How It Works - Background Indexing - Semantic Search - Composite Scoring - Troubleshooting - Common Issues - Performance Tuning - Privacy & Security

Overview: This repository is a template designed to fine-tune sentence embedding models using the `sentence-transformers` package. It supports various dataset formats and loss functions, including synthetic data generation from PDF files using Azure OpenAI. Key features: - Fine-tune any SentenceTransformer-compatible model - Supports multiple dataset formats: `positive_pair`, `triplets`, `pair_with_score` - Generates synthetic training data from PDFs using Azure OpenAI - Compatible with various loss functions: `matryoshka`, `triplet`, `contrastive`, `cosine_similarity` - Evaluates the fine-tuned model using metrics such as NDCG@k - CLI and config file-based execution - Outputs fine-tuned models and evaluation results Tech stack: Python Why it matters: This template is useful for developers looking to fine-tune sentence embedding models for various applications, including natural language processing tasks. Repository: ritesh-modi/fine-tuning-embeddings-template GitHub stars: 8 GitHub description: This repo is a template to fine-tune embedding models using the [`sentence-transformers`](https://www.sbert.net/) package. This project supports fine-tuning on various types of datasets and loss functions, including synthetic dataset generation from PDF files using Azure OpenAI. Repository: ritesh-modi/fine-tuning-embeddings-template GitHub stars: 8 GitHub description: This repo is a template to fine-tune embedding models using the [`sentence-transformers`](https://www.sbert.net/) package. This project supports fine-tuning

### Overview: PaveDB is an inspectable retrieval database designed for AI applications. It provides text-backed vector search with source provenance and query replay, making it easy to inspect and debug the retrieval process. ### Key Features: - **Source Provenance**: Every search result includes the document ID, exact snippet, and source-specific page, row, or character offset. - **Query Replay**: Replays queries to see the exact chunk that matched, allowing you to observe and compare results. - **Embedding and Retrieval**: Supports embedding and retrieving text data, with full source provenance and query replay capabilities. - **Multi-Tenant**: By default, PaveDB supports multiple tenants with tenant-specific APIs and quotas. - **Embedding Swap**: Allows swapping models per collection, ensuring compatibility with different models. - **Production-Safe**: Durable storage, transactional schema migrations, and a production-ready server. ### Tech Stack: Python ### Why It Matters: - **Inspection and Debugging**: Easily inspect and debug the retrieval process. - **Replay and Traceability**: Replay queries to observe and compare results. - **Multi-Tenant Support**: Supports multiple tenants with tenant-specific APIs and quotas. - **Embedding Swap**: Allows swapping models per collection, ensuring compatibility with different models. - **Production-Safe**: Durable storage, transactional schema migrations, and a production-ready server. ### Repository: rodrigopitanga/pavedb ### Primary Language: Python ### GitHub Stars: 8 ### GitHub Description: PaveDB is an inspect

Overview: TaskX-Nano is a lightweight Python library designed to map natural language prompts to programming functions, optimizing parameters for efficiency and resource management. Key features: - NLP-to-Function Mapping - Parameter Optimization - Lightweight Design - Extensible Function Library - Open Source Tech stack: Python Why it matters: FunctionX-Nano simplifies the integration of NLP-driven automation into workflows, offering fast and resource-efficient solutions without the need for large language models, making it ideal for developers and edge devices. Repository: rodyrahi/Taskx-nano GitHub stars: 8 GitHub description: A python library built to execute tasks with natural language