Projects

Overview: NuGuardAI/nuguard is an AI red-teaming tool and LLM security framework designed to evaluate and assess the security of agentic AI applications. It includes features for testing prompt injections, handling vulnerability assessment, generating SBOMs, and performing static analysis. Key features: - AI red-teaming tool - LLM security framework - Testing prompt injections - Vulnerability assessment - SBOM generation - Static analysis Tech stack: Python Why it matters: NuGuardAI/nuguard provides a comprehensive solution for security testing of AI applications, ensuring robust protection against potential vulnerabilities and prompt injection attacks.

### Overview: CoDA is Salesforce AI Research's open diffusion language model, providing a unified training pipeline from pre-training to post-training, evaluation, and serving. ### Key Features: - Unified training pipeline across TPU pre-training, GPU fine-tuning, evaluation, and serving. - Supports both supervised diffusion fine-tuning and end-to-end diffusion language modeling. - Includes a simple Fast-API based serving backend for deployment. ### Tech Stack: Python ### Why it Matters: CoDA offers a comprehensive solution for diffusion language modeling, making it a valuable resource for researchers and developers alike.

Overview: VisualWebBench is an evaluation framework designed to assess the understanding and grounding capabilities of multimodal language models (LLMs) in web scenarios. It consists of seven tasks and 1.5K human-curated instances from 139 real websites, covering 87 sub-domains. Key features: - Comprehensiveness: It spans 139 websites with 1.5K samples across 12 different domains and 87 sub-domains. - Multi-granularity: Assesses MLLMs at three levels: website-level, element-level, and action-level. - Multi-tasks: Includes tasks for understanding, OCR, grounding, and reasoning. - High quality: Ensured through careful human verification and curation. Tech stack: Python Why it matters: VisualWebBench serves as a valuable resource for the research community, contributing to the development of more powerful and versatile multimodal language models for web-related applications.

Overview: This repository is a collection of various tasks and resources related to machine learning and deep learning, including tutorials, books, and papers on different topics and applications. Key features: - Tutorials on Naive Bayes, Deep Learning, Text Embedding, Keyword Spotting, Voice Activity Detection, Speech Recognition, and Activity Detection. - Books and papers on machine learning and deep learning, including Scikit-learn, Naive Bayes, Deep Learning, Text Embedding, Keyword Spotting, Voice Activity Detection, Speech Recognition, and Activity Detection. - A beginners guide to understanding neural nets, detailed explanations with math, practical approaches, and code examples in Python, Theano, TensorFlow, and Keras. - A list of various tasks that can be achieved with machine learning/deep learning, including keyword spotting, voice activity detection, speech recognition, and activity detection. - A list of resources for various fields of machine learning and deep learning, including Scikit-learn, Naive Bayes, Deep Learning, Text Embedding, Keyword Spotting, Voice Activity Detection, Speech Recognition, and Activity Detection. - A list of resources for various fields of machine learning and deep learning, including Scikit-learn, Naive Bayes, Deep Learning, Text Embedding, Keyword Spotting, Voice Activity Detection, Speech Recognition, and Activity Detection. Tech stack: Python, Theano, TensorFlow, Keras. Why it matters: This repository provides a comprehensive list of resources and tutorials for various fields of machine learning and deep learning, making it an essential resource for developers and researchers looking to learn

Overview: Safe-CLIP is an enhanced vision-and-language model designed to mitigate the risks associated with NSFW (Not Safe For Work) content in AI applications. It fine-tunes the CLIP model to ensure safer outputs in text-to-image and image-to-text retrieval and generation tasks. Key features: - Safe-CLIP is an enhanced vision-and-language model. - It fine-tunes the CLIP model to mitigate NSFW content. - It includes models and datasets for both HuggingFace and open_clip. Tech stack: Python Why it matters: Safe-CLIP is crucial for ensuring ethical and safe use of AI in applications that involve sensitive content, such as removing explicit sexual content from visual and textual outputs. Repository: aimagelab/safe-clip GitHub stars: 68 GitHub description: [ECCV 2024] Official implementation of Safe-CLIP, a safety-aligned vision-language model for mitigating NSFW content in multimodal retrieval and generation. README excerpt: This repository contains the reference code for the paper [**Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models**](https://arxiv.org/abs/2311.16254). It involves explicit sexual content, racially insensitive language, and other material that may be harmful or disturbing to certain users. Please use this content solely for research purposes and proceed with caution.

Overview: This project introduces a recommender system pipeline that leverages PyTorch, Redis, Elasticsearch, Feast, Triton, and Flask for offline and online operations. The system includes offline data preprocessing, term and vector recall, and real-time inference for both offline and online scenarios. Key features: - Offline data preprocessing with PyTorch and Conda - Term and vector recall using Redis and Elasticsearch - Real-time inference with Triton and Flask Tech stack: Conda, Docker, PyTorch, Redis, Elasticsearch, Feast, Flask, Triton Why it matters: This system demonstrates the practical implementation of a recommender system, showcasing the integration of multiple technologies for efficient offline and online operations, ensuring scalability and performance.

Overview: This project provides an official API for Nano Banana, a model that generates images and previews flash images, compatible with OpenAI's API. It offers a cost-effective, token-billed endpoint that supports various resolutions and image inputs. Key features: - Token-billed pricing model - Supports multiple image resolutions - Compatible with OpenAI's API - Provides detailed pricing information Tech stack: Python Why it matters: This API allows developers to easily integrate Nano Banana's image generation capabilities into their projects without the need for a monthly subscription, making it ideal for cost-conscious applications. Repository: apimartlt3f/gemini-2.5-flash-image-preview-official-api-ja Primary language: Python GitHub stars: 68 GitHub description: Nano Banana (official) API (`gemini-2.5-flash-image-preview-official`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint. README excerpt: This README provides an overview of the Nano Banana API, including pricing, model id, and parameters. It also includes a quickstart guide for integrating the API using both Bash and Python.

Overview: Grok Imagine Image 2.0 API is a high-fidelity image generation service provided by APIMart, offering a single key to access over 300 models with a flexible pricing model. Key features: - Supports OpenAI-compatible endpoint - Pricing per image by resolution and quality - $1 minimum, pay-as-you-go - Charge comes back in the response - Async by design, tasks can be polled for status Tech stack: Python Why it matters: Grok Imagine Image 2.0 API simplifies access to a wide range of image models, providing high-fidelity results at a competitive price point. Repository: apimartlw69qg/grok-imagine-image-2.0-api-cn Primary language: Python GitHub stars: 67 GitHub description: Grok Imagine Image 2.0 API (`grok-imagine-image-2.0`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: VeriThinker is a novel approach for compressing causal reasoning (CoT) tasks. Unlike conventional methods that fine-tune large language models (LLMs) on synthetic CoT data, VeriThinker trains the models to verify the correctness of CoT solutions through an auxiliary verification task. This results in more efficient reasoning chains without the need for synthetic data, as it inherently improves the model's ability to recognize when to stop self-reflection. Key features: - VeriThinker trains LLMs to verify CoT solutions. - Reduces reasoning chain lengths while maintaining or improving accuracy. - Applied to DeepSeek-R1-Distill-Qwen-7B, it reduces reasoning tokens from 3790 to 2125 on MATH500 and from 14321 to 10287 on AIME25, with an accuracy gain of 0.8% and 2.1%, respectively. Tech stack: Python Why it matters: VeriThinker offers a more efficient and accurate way to compress causal reasoning tasks, reducing the need for synthetic data and improving model performance without overthinking.

Overview: Bedrust is a Rust application designed to invoke models on Amazon Bedrock, providing a convenient way to interact with large language models through the AWS Bedrock service. Written in Rust, it supports a variety of models including DeepSeek R1, Claude 3.7 Sonnet, and more. Key features: - Invokes models on Amazon Bedrock - Supports multiple models including DeepSeek R1, Claude 3.7 Sonnet, and others - Uses Rust for efficient and seamless integration - Configurable through a configuration file and default model selection Tech stack: Rust Why it matters: Bedrust simplifies the process of interacting with large language models on Amazon Bedrock, making it easier for developers to leverage these powerful AI tools without needing to delve into the complexities of AWS Bedrock's API directly.

Overview: Grok Imagine 2.0 is a powerful image generation and region-edit model provided through the APIMart API, offering flexible pricing and seamless integration with OpenAI-compatible endpoints. Key features: - Supports both 1K and 2K resolutions - Charges $0.015 per image, with a $0.015 discount for region-edit tasks - Compatible with multiple models and models can be switched easily - Charges are reflected in the response, and tasks are refunded for failed submissions - Async by design, tasks can be monitored via `GET /v1/tasks/{id}` Tech stack: Python Why it matters: Grok Imagine 2.0 provides a robust and flexible solution for image generation and editing, with a user-friendly pricing model and seamless integration with popular APIs, making it an ideal choice for developers looking to enhance their applications with advanced image processing capabilities.

### Overview: Blender add-on that turns Blender's viewport into a chat with **Agent OpenCode**. You describe what needs to be created or changed, and Agent OpenCode generates the corresponding Python code (`bpy`) and the add-on will execute it directly in the scene. ### Key features: - **Portable OpenCode** - downloads and runs from the add-on folder without needing to install OpenCode on the system. - **Chat in 3D Viewport** - history of messages, session agent. - **Create from Image** - attach a file or screenshot of the viewport. - **Auto-run + Auto-fix** - generated code is executed automatically; if there's an error, the agent fixes the script itself. - **Providers** - OpenAI, OpenRouter, Anthropic, Groq, DeepSeek, local Ollama. - **Security config**: agent is restricted to `bash: deny`, only local file edits allowed. ### Tech stack: Python ### Why it matters: - **Efficiency**: Quickly generate and execute Python code directly in Blender without needing to set up OpenCode on the system. - **User Experience**: Provides a seamless integration with Blender's workflow, making it easy to create or modify 3D models. - **Versatility**: Supports multiple AI providers, ensuring a wide range of capabilities. ### GitHub stars: 68 ### Repository: hitkeeen/blender-ai-addon ### Primary language: Python ### GitHub description: Blender add-on: create and edit 3D models by chatting with AI - send

### Overview: Pi Pin is an AI wearable device that records conversations and summarizes them using Generative AI. It is built on a Raspberry Pi Zero 2 W and is fully open-source, affordable, and hackable. ### Key Features: - **AI Conversation Summarization**: Records conversations and summarizes them using Generative AI. - **Affordable and Open-Source**: Built on a $15 Raspberry Pi Zero 2 W, fully open-source, and fully hackable. - **Compact Design**: Enclosure is designed around the Raspberry Pi Zero, Microphone, and battery to ensure compact physical dimensions. ### Tech Stack: Jupyter Notebook ### Why It Matters: Pi Pin is an innovative AI wearable that combines cutting-edge AI technology with affordable hardware, making it accessible and useful for anyone interested in AI and wearable technology.

Overview: PSEC is a framework designed for skill expansion and composition in the parameter space, aiming to facilitate efficient and flexible skill expansion and composition, iteratively evolve agents' capabilities, and efficiently address new challenges. Key features: - Skill Expansion: Facilitates the efficient and flexible addition of new skills to agents. - Skill Composition: Enables the composition of skills to create more complex behaviors. - Iterative Evolution: Iteratively improves the agents' capabilities through skill expansion and composition. - Efficient Challenge Solving: Quickly addresses new challenges by leveraging the evolved skills. Tech stack: Python Why it matters: PSEC provides a novel framework that addresses the challenges of skill expansion and composition in the parameter space, making it an essential tool for researchers and developers looking to enhance the capabilities of their agents in dynamic and evolving environments.

Overview: This project provides a fine-tuned streaming speech recognition model for Persian (Farsi) using NVIDIA's Nemotron 3.5 ASR streaming technology. The model is trained on 1,181 hours of publicly licensed Persian speech and evaluates low word error rates across various test sets. Key features: - Fine-tuned Nemotron 3.5 ASR streaming model - Streaming transcription of live audio and file uploads - Supports multiple test sets including FLEURS, Common Voice, and held-out conversational test sets - Lower latency with a 1.12 s look-ahead - Uses the 🤗 Transformers version of the model Tech stack: Python Why it matters: This project offers a practical tool for real-time Persian speech recognition, which can be crucial in applications such as voice assistants, language translation, and automated transcription. Repository: [mallahyari/nemotron-asr-streaming-farsi](https://github.com/mallahyari/nemotron-asr-streaming-farsi) GitHub stars: 66 GitHub description: Streaming Persian (Farsi) speech recognition: fine-tuning NVIDIA Nemotron 3.5 ASR streaming. Data prep, training, evaluation, inference.

### Overview: Multimodal Graph Learning (MMGL) is a systematic framework designed to capture and utilize information from multiple multimodal neighbors within complex, interconnected data structures. This approach aims to enhance language models (LLMs) by augmenting their text generation capabilities with contextual information from related entities. ### Key features: - **MMGL Framework**: A comprehensive framework for multimodal graph learning that integrates text, image, and other modalities. - **Generative Tasks**: Specifically optimized for generative tasks, enhancing language models' ability to generate text based on multimodal inputs. - **Pretrained Models**: Leverages pretrained language models (LMs) to augment text generation with contextual information from multimodal neighbors. ### Tech stack: The project is implemented using PyTorch, a popular deep learning framework. ### Why it matters: By integrating multimodal information, MMGL aims to improve the performance of language models in a variety of tasks, including text generation, by providing richer, more contextually relevant input data. This could lead to more accurate and diverse outputs from LLMs, making them more useful in real-world applications.

### Overview: ColorPeel is a project that aims to generate geometric shapes with target colors, thereby disentangling color and shape concepts. By learning on multiple color-shape images, the method successfully separates color and shape, enabling personalized color usage in image generation. ### Key Features: - **Color and Shape Disentanglement**: ColorPeel generates shapes with specific target colors, effectively separating color and shape concepts. - **Diffusion Models**: Utilizes diffusion models to learn on multiple color-shape images, improving disentanglement. - **Custom Datasets**: Offers two sample datasets for learning colors from 2D and 3D shapes, and allows users to create their own 2D dataset using a provided script. ### Tech Stack: Python ### Why It Matters: ColorPeel is significant because it facilitates the disentanglement of color and shape concepts, making it easier to personalize color usage in image generation. This project demonstrates the potential of using diffusion models for learning and disentangling complex image features. ### GitHub Stars: 68 ### Repository: moatifbutt/color-peel ### Primary Language: Python ### Future Work: While experimenting with various color mappings, the project encountered convergence issues that are currently unresolved. Future research directions include developing training schemes based on Textual Inversion for mapping color spaces into text embeddings.

Overview: `flux.swift.cli` is a CLI tool designed for generating images based on text prompts using the FLUX.1 model, offering customizable parameters and support for multiple models. Key features: - Text-to-image generation - Image-to-image generation using both text prompts and reference images - Customizable parameters for image size, inference steps, guidance scale, and more - Model optimization options for quantization and float16 precision - Support for multiple FLUX.1 models: Schnell, Dev, and Kontext Dev Tech stack: Swift Why it matters: This tool simplifies the process of generating images from text prompts, making it accessible for developers and enthusiasts alike.

Overview: Este projeto é um Single Page Application (SPA) desenvolvido com o Framework Quasar 1.0 e utiliza a API de Reconhecimento de Fala da Web para capturar áudio e transformá-lo em texto, ou vice-versa. Key features: - Captura de áudio - Transformação de áudio em texto - Utilização de texto como base para emitir áudio Tech stack: JavaScript, Quasar Framework 1.0, Speech API Why it matters: Este projeto oferece uma solução prática para integrar reconhecimento de voz e áudio em aplicações web, facilitando a interação com o usuário de maneira mais natural e intuitiva.

Overview: The Pinecone Go SDK is a Go client designed for interacting with Pinecone Database, providing full-text search, vector search, and more. Key features: - Documents API: Full-text search, query strings, and vector search over indexes with a document schema. - Integrated embedding: Embed document text with hosted models for meaning-based searches. - Vectors API: Upsert, query, fetch, update, delete, and list records in vector indexes. - Legacy Records API: Upsert and search text in legacy integrated indexes. - Index management: Create, configure, and delete indexes, and manage backups, collections, and namespaces. - Inference API: Generate embeddings and rerank results with models hosted by Pinecone. - Admin API: Manage projects, organizations, API keys, service accounts, role bindings, invites, and users. - Retries: Optional automatic retries with exponential backoff. Tech stack: Go Why it matters: The Pinecone Go SDK simplifies interaction with Pinecone Database, making it easier to manage indexes, search documents, and integrate with Pinecone's powerful models and APIs. Repository: pinecone.io/go-pinecone GitHub stars: 68