Projects

Overview: Speech to speech bot built with Python, supporting natural language processing and integrating with various libraries. Key features: - Recognizes speech input - Looks for an answer using Chatterbot - Replies with speech synthesized by Google Text-to-Speech Tech stack: Python, Google Text-to-Speech, SpeechRecognition, Chatterbot, PyGame Why it matters: This project offers a versatile solution for creating speech-based AI applications, making it useful for video game integration and smart home automation.

Overview: This repository contains the code for the AI Builders Bootcamp #1, a comprehensive course on AI development and machine learning. Key features: - Introduction to AI and basic automation scripts with Python - ETL pipelines for machine learning projects - Building AI systems with LLMs (Large Language Models) - Analyzing unstructured survey results - Semantic search over blog posts - Fine-tuning LLMs for specific tasks - Project management framework for AI projects Tech stack: Python, Jupyter Notebook, GitHub Why it matters: This course provides a structured approach to learning AI and machine learning, from basic automation to advanced models, making it a valuable resource for developers looking to build and deploy AI solutions.

Overview: TagForge is an advanced AI tagging and prompt generation studio that leverages modern LLMs to craft perfect Stable Diffusion prompts in a high-performance desktop environment. Key features: - Multi-Model Tagging: Specialized generator for Stable Diffusion and FLUX/Midjourney. - Multimodal Vision: Precise metadata extraction from images or high-fidelity captions. - Contextual Chat: Persistent, coding-capable assistant with real-time access to provider-specific features. Tech stack: C# and .NET 9. Why it matters: TagForge provides a sleek, high-performance workspace for AI prompting and tagging, ensuring perfect Stable Diffusion prompts and precise image descriptions.

### Overview: ClawWatch is the first intelligent AI agent running natively on a smartwatch, utilizing NullClaw for offline speech recognition and TTS integration. It offers a voice-controlled smartwatch assistant with no cloud dependency, ensuring fast and efficient responses. ### Key Features: - **NullClaw Integration**: Leveraging NullClaw for offline speech recognition and TTS. - **Real-Time Connectivity**: Employs offline speech recognition and TTS for real-time interaction with other agents and the watch's sensors. - **Vector Graphics Avatars**: New avatars enhance the watch's character and connection to real-time discussions. - **Live Body Sensing**: Real-time connection to watch sensors for speaking about vital signs and movements. ### Tech Stack: Kotlin ### Why It Matters: ClawWatch represents a significant advancement in smartwatch technology, combining AI with native capabilities to create a truly integrated and interactive experience. Its real-time connection to both agents and the watch's sensors makes it exceptionally versatile and useful for monitoring and interacting with the user's environment.

### Overview: Nano Banana Pro prompts offer tested image prompt recipes, including example renders, aspect ratios, and the model ID (`gemini-3-pro-image-preview`). This repository provides an API key to access Nano Banana Pro and includes cost per image and pricing information. ### Key Features: - **Model ID:** `gemini-3-pro-image-preview` - **Recipes:** 8 tested prompts (machine-readable in `data/prompts.json`, table in `PROMPTS.md`) - **Effective Price:** $0.03 per image (estimated cost of the set: $0.24) ### Tech Stack: - OpenAI-compatible endpoint ### Why It Matters: - **Cost Efficiency:** Nano Banana Pro is priced at $0.03 per image, making it highly cost-effective. - **Minimal Setup:** No need for prepayment or subscription plans. - **Clear Pricing:** All operations are charged based on the actual cost, with refunds for failed tasks. - **Asynchronous Design:** Tasks are completed by polling the `/v1/tasks/{id}` endpoint, eliminating the need for batch and retry integration. ### Other Names: - **Nano Banana Pro** - **Nano Barana** ### Recipes Overview: - **Product / E-commerce:** Frosted glass perfume bottle on a wet black stone slab. - **Food Photography:** Spicy miso ramen bowl with soft boiled egg and nori. - **Cinematic Still:** Rain-soaked Kyoto alley at night, paper lantern reflections on wet stone. ### Prompt Structure Cheat-Sheet: | Slot |

### Overview: This demonstration showcases the use of Natural Language Query (NLQ) to interact with an Amazon RDS for PostgreSQL database, leveraging various AI models and tools such as Amazon SageMaker JumpStart, Amazon Bedrock, LangChain, Streamlit, and Chroma. The application's web-based interface, powered by LangChain, Streamlit, and Chroma, allows users to ask natural language questions and receives natural language answers along with SQL queries and Pandas DataFrame-compatible results. ### Key Features: - **Foundation Model Choice and Accuracy of NLQ**: The selection of the Foundation Model (FM) for NLQ is crucial. OpenAI's GPT-3 and GPT-4 series models, Amazon Titan Text G1, and open-source models like `google/flan-t5-xxl-fp16` were tested. Amazon Titan Text G1 provided accurate responses to basic queries, while open-source models could answer some simple queries but struggled with more complex ones. ### Tech Stack: - **LangChain**: A library for building generative AI applications. - **Streamlit**: A Python-based web application framework. - **Chroma**: A vector database for storing and retrieving text data. - **Amazon SageMaker JumpStart**: A platform for deploying and managing machine learning models. - **Amazon Bedrock**: A managed AI and ML infrastructure service. - **Amazon RDS for PostgreSQL**: A relational database service. ### Why It Matters: This demonstration highlights the importance of choosing the right Foundation Model and optimizing NLQ accuracy through prompt engineering and labeled sample queries.

### Overview: Model serving is a crucial process in deploying machine learning models. It involves deploying models to different environments based on their performance requirements. High-volume predictions often require batch processing, while low-latency needs are best met with real-time endpoints. This project collection includes various libraries and techniques for serving models, such as TensorFlow Serving, TorchServe, BentoML, Clipper, and more, providing a comprehensive solution for different use cases. ### Key Features: - **Tensorflow Serving**: High-performant framework for serving Tensorflow models via gRPC protocol. - **TorchServe**: Flexible and easy-to-use tool for serving PyTorch models. - **BentoML**: Open source framework for high-performance ML model serving. - **Clipper**: Model server project from Berkeley's Rise Rise Lab, supporting TensorFlow, Scikit-learn, and Caffe models. - **Cortex**: Open-source platform for deploying machine learning models as production web services. - **Multi-Model-server**: Flexible and easy-to-use tool for serving deep learning models. - **DeepDetect**: Machine Learning production server for TensorFlow, XGBoost, and Caffe models. - **Aml-Deploy**: GitHub Action for deploying Machine Learning Models to Azure. - **MLOps**: Empowers data scientists and app developers to bring ML models to production. - **ForestFlow**: Cloud-native machine learning model server. - **Jina**: Cloud native search framework for deep learning models. - **KFServing**: Serverless framework for deploying and monitoring machine learning models in Kubernetes. - **NVIDIA

### Overview: This repository is a comprehensive collection of AI APIs designed for immediate product development. It includes a wide range of APIs such as language models, computer vision, agents, and more, all ready for production use. With 1,555 APIs across 5 categories, it's a go-to resource for developers looking to ship real products quickly. ### Key Features: - **LLMs (Language Models):** Tools for language generation and text analysis. - **Computer Vision:** APIs for image and video processing. - **Agents:** Tools for creating conversational agents. - **MCP Servers:** Servers for machine learning models. - **APIs:** A variety of APIs for different functionalities. ### Tech Stack: The tech stack is designed to be efficient and user-friendly, making it easy for developers to integrate and use the APIs. ### Why It Matters: This repository stands out due to its comprehensive coverage of AI APIs, making it a valuable resource for developers looking to quickly ship real products. The APIs are production-ready, ensuring they can be integrated seamlessly into existing projects.

Overview: This is a Rust SDK for the Deepgram API, designed to simplify integration with their services. Key features: - Speech-to-text: Pre-recorded REST, live WebSocket streaming, Flux STT - Text-to-speech: Aura REST and streaming WebSocket, Flux TTS REST and streaming WebSocket - Text Intelligence: Sentiment, summary, topics, intents - Management API: Projects, keys, members, usage, billing, models - Voice Agent WebSocket client: /v1/agent/converse Tech stack: Rust Why it matters: This SDK provides a straightforward interface for developers to integrate with Deepgram's powerful speech processing and AI capabilities. Repository: deepgram/deepgram-rust-sdk GitHub stars: 68 GitHub description: Community Rust SDK for Deepgram.

Overview: This project implements the "Denoising Diffusion Probabilistic Models" paper in PyTorch, providing a practical tool for generating images through diffusion processes. Key features: - Diffusion process implementation - Real-time image generation - Versatile parameter tuning - Compatibility with PyTorch Tech stack: PyTorch Why it matters: This project offers a practical and accessible way to experiment with and understand the theoretical concepts behind diffusion models, making it valuable for developers and researchers alike.

Overview: This project generates images using powerful text-to-image models available through Cloudflare Workers AI. Key features: - Generate images - Use Cloudflare Workers AI models - Download generated images Tech stack: TypeScript Why it matters: This project simplifies the process of generating images using AI, making it accessible and easy to use for developers and designers alike.

Overview: This repository contains the implementation of the paper "Deep Multimodal Subspace Clustering Networks" by Mahdi Abavisani and Vishal M. Patel. The paper investigates various fusion methods for multimodal subspace clustering and proposes a new fusion technique called "affinity fusion." Key features: - Tensorflow implementation - Pretrained model for demonstration - Data preprocessing steps Tech stack: Python Why it matters: This repository provides a practical implementation of a novel method for multimodal subspace clustering, offering insights into effective fusion techniques for data integration in deep learning applications.

### Overview: Personal AI Memory is a local-first Chrome extension that captures and stores conversations from various AI platforms like ChatGPT, Claude, Gemini, Perplexity, and Grok. It leverages in-browser Hybrid RAG (Vector + BM25) and semantic search to provide relevant context in new chats. The extension is 100% private, with no servers or API keys required, ensuring complete privacy and security. ### Key features: - **Captures Conversations:** Silently captures and stores conversations from various AI platforms. - **Local Storage:** Stores conversations locally, indexed for quick access. - **Semantic Search:** Provides semantic search capabilities to enhance recall. - **Recall Button:** A one-click recall button to inject relevant context into new chats. - **100% Privacy:** No servers, no cloud storage, no account required. - **No API Keys:** No need for API keys or external services. - **Chrome Web Store Integration:** Easy installation via Chrome Web Store or manual download. - **Development Mode:** Auto-rebuilds on every save for easy development and testing. ### Tech Stack: - TypeScript - WebAssembly - IndexedDB ### Why it Matters: Personal AI Memory is designed to help users remember and recall conversations from various AI platforms, ensuring that they can quickly access relevant context in new chats. The local-first approach ensures complete privacy and security, making it an ideal solution for users who value their privacy and want to avoid the cloud.

### Overview: ASRT SDK WinClient 是一个用于 ASRT 语音识别系统的 Windows 客户端 SDK 和演示客户端软件。 ### Key features: - Windows 客户端 SDK - Demo 客户端 - 支持 ASRT 语音识别系统 API 调用 - 实现语音识别应用开发的演示环境 ### Tech stack: C# .NET Framework 4.7 ### Why it matters: 适用于 ASRT 语音识别系统的客户端工具，为开发者提供一个演示环境，便于进一步开发语音识别应用软件。

Overview: Rustic UI Components is a React component library designed to simplify the creation of conversational user interfaces, focusing on usability and integration. Key features: - A wide range of reusable components for chatbot and virtual assistant development. - Support for various chat functionalities and multimodal interactions. - Available on Figma for design exploration. Tech stack: JavaScript, React. Why it matters: Rustic UI Components provides essential building blocks for developers to create engaging and dynamic chat-based user experiences, making it a valuable tool for building conversational applications. Repository: rustic-ai/rustic-ui-components GitHub stars: 71 GitHub description: React component library for crafting user-friendly and engaging conversational experiences. Figma file: [Visit Figma](https://www.figma.com/community/file/1340378790384732080) Storybook: [Visit Storybook](https://rustic-ai.github.io/rustic-ui-components) CONTRIBUTING.md: [Building From Source](CONTRIBUTING.md#building-from-source-and-using-it-locally) LICENSE: MIT License

### Overview: Enhance your language learning with our Pronunciation Assessment API. Tailored to support language learners, our API evaluates pronunciation by analyzing audio files and text strings, providing comprehensive metrics for speech quality and fluency. ### Key Features: - **Pronunciation Score (Phonemes, Words, Sentences)** - **Fluency Score** - **Completeness Score** - **Rhythm Score or Speed** - **Syllable Stress Analysis** - **Liaison Detection** - **Loss of Plosion Identification** - **Rising/Falling Tone Analysis for Sentence Endings** - **Real-Time Feedback at the Sentence Level** - **Phoneme-Level Mispronunciations (Insertion, Deletion, Substitution)** ### Tech Stack: C# ### Why It Matters: Elevate your language proficiency with precise analysis and instant feedback mechanisms. Access robust tools to evaluate and refine your pronunciation skills within structured contexts.

Overview: This repository contains a collection of deep learning models for remote sensing spatiotemporal fusion, including CNN, Transformer, GAN, Diffusion Model, Sequence Model, and others. It serves as a reference for researchers and practitioners in the field. The repository aims to provide a comprehensive overview of the advancements, challenges, and opportunities in remote sensing spatiotemporal fusion. Key features: - Deep learning models for remote sensing spatiotemporal fusion - Relevant survey papers - Categories of deep learning models - Commonly used datasets in the field - Code implementation for deep spatiotemporal fusion models Tech stack: Python Why it matters: This repository provides a valuable resource for researchers and practitioners interested in remote sensing spatiotemporal fusion, offering insights into the latest advancements, challenges, and opportunities in the field. GitHub stars: 70 Repository: yc-cui/Deep-Learning-Spatiotemporal-Fusion-Survey

### Overview: This project aims to develop a multilingual speech-to-text translation system using deep learning techniques. The system will utilize a hybrid model combining Gaussian Mixture Models (GMM) and Neural Networks (NNet) to accurately translate multilingual audio into text. ### Key Features: - **Multilingual Support**: The system will handle both Chinese and English in the same sentence. - **Deep Learning Architecture**: Utilizes DNN, CNN, and LSTM to improve translation accuracy. - **Model Integration**: Combines GMM and NNet for enhanced performance. ### Tech Stack: - **GitHub Repository**: yh1008/speech-to-text - **Primary Language**: Jupyter Notebook - **GitHub Stars**: 71 ### Why it Matters: This project is crucial for multilingual communication applications, particularly in the context of language learning and cross-cultural interactions. It addresses a significant gap in current speech recognition technology, providing a practical solution for multilingual users. ### Repository: - **Primary Language**: Jupyter Notebook - **GitHub Description**: A mixlingual speech recognition system using deep learning architecture (DNN, CNN, LSTM) for multilingual audio translation. ### Required Packages: - **Run Kaldi on GPU**: Ensure GPU support for Kaldi operations. - **Install Tensorflow-gpu**: Install the GPU version of TensorFlow. - **Install Theano-gpu**: Install the GPU version of Theano. - **Keras-kaldi**: Use Keras with Kaldi for efficient model training. ### Directory Description: - **

Overview: This project is an optimized version of llama.cpp specifically designed for NVIDIA Tesla V100 (SM70) GPUs, achieving high throughput and low memory usage. It supports Qwen3.8-Flash-Next with MTP and Qwen3.8-27B models, enabling concurrent serving and multimodal processing. Key features: - Optimized for NVIDIA V100 (SM70) hardware - Supports concurrent serving and multi-modal processing - High throughput and low memory usage - Compatible with 2-6 GPUs, with 262K-512K context - Built-in web UI for easy deployment and management Tech stack: C++ Why it matters: This project is crucial for developers looking to optimize their AI models for high-performance GPU environments, particularly for tasks that require both fast processing and efficient memory management. Repository: 1115714829/llama.cpp-v100 GitHub stars: 62 GitHub description: llama.cpp optimized for NVIDIA Tesla V100 (SM70): Qwen3.8-Flash-Next with MTP (prefill 3402 tok/s at 200K context on 6x V100-16GB; also runs on 2x V100-16GB) and Qwen3.8-27B with DFlash2 (decode up to 220 tok/s); 2-6 GPUs, 262K-5

### Overview MODUS is a unified multimodal generation model that can transform any input modality into any output modality using a single causal transformer. This approach eliminates the need for separate encoders, decoders, and task-specific pipelines, making it efficient and flexible. ### Key Features - **Unified Modality Handling:** MODUS uses a single causal transformer trunk to process any input modality and any output modality. - **Cross-Entropy for Text:** The 1D expert handles discrete sequences like text and grounding boxes with autoregressive next-token prediction. - **Flow Matching for Spatial Latents:** The 2D expert matches continuous spatial latents using flow matching. - **Shared Trunk:** Every modality goes through the same trunk, ensuring consistency and avoiding modality-specific weights. ### Tech Stack - **Python:** The primary programming language used for the implementation. ### Why It Matters MODUS simplifies multimodal generation by removing the need for separate encoders, decoders, and task-specific pipelines, making it more efficient and flexible. Its unified approach to multimodal generation is particularly useful in scenarios where modality-specific models are not feasible or desirable. ### Repository - **Repository:** EPFL-VILAB/Modus - **GitHub Stars:** 70 - **GitHub Description:** MODUS: Decoder-only Any-to-Any Modeling of Diverse Modalities [ICML 2026] ### Updates - **2026/10:** Added the optional **MODUS-Hunyuan** inference backend and checkpoint.