Projects

Overview: Gemini Nano Banana 2.1 API is an OpenAI-compatible endpoint for Gemini Nano Banana 2.1, offering image generation and reference-image editing at $0.0137 per image. It supports various resolutions and is compatible with 300+ models. Key features: - Supports text-to-image and reference-image editing - Per-image pricing with tiered resolution options - Compatible with 300+ models and 1K, 2K, and 4K outputs - Asynchronous processing with task ID polling - Real-time billing and cost reporting Tech stack: Python Why it matters: Gemini Nano Banana 2.1 leverages APIMart's OpenAI-compatible endpoint, providing a cost-effective solution for image generation and editing without the need for a monthly subscription. Repository: a3g9df6h4jh/nano-banana-2-1-api-id Primary language: Python GitHub stars: 47 GitHub description: Gemini Nano Banana 2.1 API (`gemini-nano-banana-2.1`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: This project provides a Nano Banana Pro (official) API endpoint for image previews, compatible with OpenAI's API, with support for text and image inputs and outputs. The API is billed per million tokens, making it cost-effective for users. Key features: - OpenAI-compatible API endpoint - Supports text and image inputs and outputs - Token-billed pricing model - Supports 1:1 aspect ratio and multiple image outputs - Supports reference image URLs and base64 data URLs Tech stack: Python Why it matters: This API is designed for developers and teams looking to integrate image generation capabilities into their applications without the need for a monthly subscription, making it ideal for cost-sensitive projects or those with unpredictable usage patterns. Repository: apimartbv/gemini-3-pro-image-preview-official-api-cn GitHub stars: 73 GitHub description: Nano Banana Pro (official) API (`gemini-3-pro-image-preview-official`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: This project is an official API for the Nano Banana Pro model, providing access to OpenAI-compatible endpoints via APIMart, including pricing, parameters, and example usage. Key features: - Access to the Nano Banana Pro model via OpenAI-compatible API - Pricing based on token usage, with a minimum charge of $1 per million tokens - Support for both text and image generation - Flexible resolution options - Real-time billing and refund on failed tasks Tech stack: Python Why it matters: This API provides a cost-effective and scalable solution for developers looking to integrate image generation into their applications, leveraging the capabilities of the Nano Banana Pro model without the need for a monthly subscription.

Overview: EmbodiedAgents is a fully-loaded ROS2-based framework designed for creating interactive physical agents that can understand, remember, and act upon contextual information from their environment. Key features: - Agents that adapt on the fly, reacting to events and component outputs. - Classical and learned manipulation, with MoveIt and VLA. - Run fully offline, with built-in local models for LLM, VLM, detection, speech-to-text, and text-to-speech. Tech stack: Python Why it matters: EmbodiedAgents provides a robust and flexible framework for deploying physical AI agents, enabling them to interact with their environment in a way that mimics human-like intelligence and decision-making capabilities.

### Overview: This repository contains a comprehensive compilation of LLM papers presented at the International Conference on Learning Representations (ICLR) in 2024. ### Key features: - **MIN-K PROB**: A method for detecting pretraining data used to train large language models. - **AgentBench**: A benchmark for evaluating LLMs as agents in interactive environments. ### Tech stack: Unknown ### Why it matters: This repository provides valuable insights into the latest advancements in large language model research, highlighting the importance of evaluating LLMs as agents in real-world scenarios.

### Overview: AI-ML Companion is an interactive platform designed to teach AI and machine learning by showing concepts in action. It includes 28 tracks and 400+ modules, covering Python, math, deep learning, LLMs, agentic AI, MLOps, and cloud GenAI, with 23 end-to-end projects and a free tier of 8 tracks. ### Key Features: - **Interactive Visualizations**: Live, animated visualizations for concepts. - **Runnable Python**: Code in the browser for hands-on experimentation. - **Narrative Narration**: Trilingual (English, Hindi, Spanish) with technical terms in English. - **PWA (Progressive Web App)**: Works on both phone and desktop, installable and offline-capable. - **Free Tier**: 8 complete tracks plus the first module of every Premium track. ### Tech Stack: Jupyter Notebook ### Why it Matters: It provides a practical, intuitive way to learn AI and ML by showing how concepts work in real-world applications, rather than just theoretical explanations.

### Overview: This project involves the development of a deep convolutional neural network (DCNN) model for the classification of skin lesions using transfer learning techniques. The model utilizes pre-trained networks such as VGG16, InceptionV3, Inception-ResNet V2, and DenseNet 201, fine-tuned on a dataset of dermatoscopic images of skin lesions. ### Key Features: - **Model Architecture**: The project focuses on fine-tuning the last convolutional block of VGG16, the top 2 inception blocks of InceptionV3, the Inception-ResNet-C of Inception-ResNet V2, the last dense block of DenseNet 201, and all layers of pre-trained Inception V3 and DenseNet 201. - **Transfer Learning**: The model leverages pre-trained networks to initialize weights, which are then fine-tuned on the specific dataset of dermatoscopic images. - **Performance Metrics**: The project includes baseline models and fine-tuned versions of VGG16, InceptionV3, Inception-ResNet V2, DenseNet 201, and Inception V3, with performance metrics showing improvements over the baseline. ### Tech Stack: - **Keras 2.2.4** - **TensorFlow 1.11** - **Batch Normalization (BN)**: Fine-tuning the BN layers requires careful handling to ensure consistent performance during inference and training. ### Why It Matters: The project demonstrates the effectiveness of transfer learning in improving

### Overview: This repository supports communities in building voice datasets for language preservation and AI development. It is a companion to the book "Envisioning Indigenous Methods in Digital Media and Ecologies," focusing on AI techniques for indigenous cultural expression. ### Key Features: - **Community Agreement Template**: A template for speaker consent with ownership tiers and withdrawal rights. - **Indigenous Datasheet**: Institutional governance framework based on OCAP® for the dataset. - **What Not to Digitize**: Decision framework for recordings that should not become training data. - **What Not to Train On**: Extends the framework to images, text, and cultural artifacts. - **Recording and Segmenting**: Record audio and split on silence. - **SNR Analysis**: Signal-to-noise quality gate. - **Transcription**: Transcribe English, Spanish, Māori, and other languages using Whisper, MMS, or manual methods. - **Augmentation**: Speed, pitch, and noise variations of consented recordings. - **Export LJSpeech**: Filter by consent tier and export in LJSpeech format. - **Export Metadata**: Merge markers with CARE metadata template. - **Scripts**: Scripts for recording, segmenting, transcribing, and exporting data. ### Tech Stack: Jupyter Notebook ### Why It Matters: This repository provides a structured approach for communities to create and manage voice datasets, ensuring ethical and effective use of data in AI development, language preservation, and cultural expression.

Overview: This Python package provides comprehensive error rates for Korean STT output, including CER, WER, CRR, and keyword/phrase preservation metrics, comparing cloud-based and open-source tools like Microsoft Azure Speech, Amazon Transcribe, Google Cloud Speech-to-Text, and OpenAI Whisper. Key features: - Comprehensive error rate analysis for STT outputs - Support for various STT systems and models - Detailed comparison of multiple STT systems - Output in JSON and Markdown formats for reproducibility and easy comparison Tech stack: Python Why it matters: This package offers a reliable and detailed way to evaluate and compare speech-to-text systems, making it valuable for developers and researchers in natural language processing and speech technology. Repository: hyeonsangjeon/computing-Korean-STT-error-rates Primary language: Python GitHub stars: 73

Overview: This repository contains reproducible LLM benchmarks for Apple Silicon devices, including Qwen 3, Qwen 0.6B, MLX, LiteRT-LM, and Core ML models. Key features: - Reproducible benchmarks for Apple Silicon devices - Quantization and capture sessions for various models - Metrics include decode tok/s, batch 1, and session details Tech stack: Python Why it matters: This benchmark provides a standardized way to measure and compare the performance of different LLM models on Apple Silicon devices, ensuring consistent results across different runs and sessions.

Overview: Simplify large ONNX models, suitable for models larger than 2GB. Key features: - Simplify large ONNX models exported from Stable Diffusion or large language models. - Supports command-line interface for easy execution. - Option to specify output file path or directory. - Option to set input shape for ONNX models. - Option to disable optimizations in ONNXSim. Tech stack: Python Why it matters: This project is essential for developers working with large models, ensuring efficient model management and optimization for deployment and inference.

### Overview: Welcome to the University of Tehran's Resources Repository! This curated list of valuable resources is dedicated to supporting students and the academic community. ### Key Features: - **Graduate Courses**: - **Trustworthy AI** - Dr. Mostafa Tavassolipour - **NLP** - Dr. Hesham Faili - **Data Mining** - Dr. Azade Shakeri - **Information Retrieval** - Dr. Azade Shakeri - **Social Network** - Dr. Masoud Asadpour - **Advanced Deep Learning** - Dr. Ahmad Kalhor - **Reinforcement Learning** - Dr. Majid Nili Ahmad Abadi - **Machine Learning** - Dr. A. Dehaqani, Dr. Mostafa Tavassolipour - **Deep Learning (with applications in vision and voice)** - Dr. Reshad Hosseini - **Statistical Inference** - Dr. Behnam Bahrak - **Deep Generative Model** - Dr. Mostafa Tavassolipour - **Neural Networks and Deep Learning** - Dr. Ahmad Kalhor - **Cognitive Science** - **Information Technology Management (ITM)** - Dr. Fattaneh Taghiyareh - **Data Analysis** - **Advanced Optimization** - **Crypto Currencies** - **Distributed Optimization and Learning** - **Bl

### Overview: Suflyor is a free, open-source teleprompter app designed for Android that follows your voice. It scrolls the script as you speak, making it perfect for creators who want to read captions for Instagram, TikTok, and camera apps. The app works offline and has no internet permission, ensuring a seamless experience. ### Key Features: - **Voice-Activated Teleprompter**: Floats over Instagram, TikTok, and camera apps, scrolling the script as you speak. - **Script Library**: Import scripts from various formats, and preview them. - **Script Editor**: Edit and modify scripts easily. - **Settings**: Customize text size, speed, and remote controls. - **Language Support**: Supports English and Russian, with the app picking the correct language based on the script's text. ### Tech Stack: - Kotlin - Speech Recognition: Sherpa-Onnx ### Why It Matters: Suflyor is perfect for creators who want to read captions for their videos without having to chase the text. It works offline, ensuring a consistent experience, and supports multiple languages, making it accessible to a wider audience. ### Repository: Repository: **olerast67/voice-teleprompter-android** GitHub Stars: **62** GitHub Description: Suflyor (Суфлёр) — free voice-following teleprompter app for Android. Scrolls as you speak, floats over Instagram, TikTok and the camera. English and Russian speech recognized on the phone, offline, no internet permission.

### Overview: Context Guru is a developer-focused tool designed to optimize AI coding agents by reducing token costs and latency without compromising accuracy. It offers two primary strategies: minimizing redundant context and keeping prompt caches warm to lower volatile costs. ### Key Features: - **Token Cost Reduction**: Streamlines context to lower token costs. - **Latency Reduction**: Delivers faster response times by optimizing prompt sizes. - **Prompt Cache Management**: Keeps prompt caches warm to reduce volatile costs. ### Tech Stack: Go ### Why It Matters: Context Guru provides a straightforward and effective solution for reducing the cost of AI model traffic while maintaining performance. It offers a simple, out-of-the-box approach to optimizing agent traffic without requiring complex configurations or opt-ins. This makes it an ideal choice for developers looking to improve their AI coding agent's efficiency and cost-effectiveness.

### Overview: This repository contains Jupyter notebooks for the courses "Hugging Face in 4 Hours" by Sinan Ozdemir, published by Pearson. The course covers effective best practices and industry case studies in using Large Language Models (LLMs) from Hugging Face. ### Key features: - Interactive learning and practical application. - Introduction to Hugging Face. - Prototyping with Hugging Face. - Fine-tuning BERT for classification. - Introduction to SmolAgents. - Multimodal models using HF. - Advanced fine-tuning of Llama 3. ### Tech stack: Jupyter Notebook ### Why it matters: This course provides an inclusive curriculum for non-developers and developers alike, offering skills to fine-tune pre-trained models for various NLP tasks and understanding how to deploy models with efficiency.

### Overview: Lycoris is a comprehensive macOS app designed for real-time speech recognition and AI-powered note-taking. It offers offline and online modes, supports multilingual transcription, and includes Japanese translation. The app is optimized for Japanese users and aims to support transcription and translation in 19 languages in the future. ### Key features: - **Real-time Speech Recognition**: - **Offline Mode**: - **Vosk for Real-time Transcription**: Uses Vosk for initial real-time transcriptions, which are then refined using Whisper for enhanced accuracy. - **Whisper for Detailed Transcription**: The primary transcription model for offline mode, Whisper delivers high-quality, multi-language transcription. For Japanese-only use cases, Kotoba-Whisper is available as an alternative for optimized Japanese transcription. - **Online Mode**: - **OpenAI Whisper API**: Integrates with the OpenAI Whisper API for high-quality transcriptions with cloud resources. - **AmiVoice API**: Additional support for AmiVoice API for even better results. - **AI-Powered Note-Taking**: - **Transcription and Translation**: Utilizes AI to transcribe and translate speech into text, enhancing the note-taking experience. - **Japanese Translation**: Supports Japanese language for users, offering optimized Japanese transcription using Kotoba-Whisper. ### Tech stack: - TypeScript ### Why it matters: Lycoris is designed to enhance your note-taking experience with real-time speech recognition, transcription, and translation capabilities. It strikes a balance between flexibility and privacy, making it

Overview: This repository contains the official PyTorch implementation for two works: Stream.FM and MelFlow, focusing on real-time streamable generative speech restoration with flow matching. Key features: - Stream.FM: Real-Time Streamable Generative Speech Restoration with Flow Matching - MelFlow: Real-Time Streaming Mel Vocoding with Generative Flow Matching Tech stack: Python Why it matters: This repository provides a comprehensive solution for real-time speech restoration tasks, leveraging generative flow matching techniques to achieve high-quality audio restoration in real-time. GitHub stars: 72 Repository: sp-uhh/streamfm

Overview: TBSM, or Three-Body Scattering Model, is a generative modeling technique that aims to generate one-step images directly from text input, bypassing traditional adversarial or diffusion models. It achieves this by leveraging three-body interactions to create a more efficient and direct method for image generation. Key features: TBSM simplifies the generative process by reducing the need for adversarial critics, teacher queries, or batch-wide pairwise fields. It achieves this by mapping a proper distributional energy directly into sample-level supervision, eliminating the need for additional training steps or stabilization tricks. This results in a more straightforward and efficient model that can generate images in a single step. Tech stack: TBSM is implemented using Python, leveraging libraries such as TensorFlow and PyTorch for its computational needs. The model architecture is designed to be lightweight and efficient, making it suitable for real-time applications. Why it matters: TBSM offers a significant improvement in generative modeling by providing a more direct and efficient approach to image generation. By eliminating the need for complex training pipelines and adversarial objectives, TBSM can achieve comparable results with fewer computational resources, making it a promising tool for applications requiring fast and accurate image generation from text inputs.

Overview: Convert your ebooks to audiobooks with speech synthesis. Key features: - Convert spoken notes from eBooks - Generate audiobooks from ebooks - Convert single images using a script Tech stack: Python Why it matters: This project simplifies the process of turning text from lectures or books into audio, making it easier to listen to and enjoy reading.

Overview: The Pulse of Motion project measures the physical frame rate of videos, providing a more accurate temporal scale than traditional methods. It addresses issues with state-of-the-art video generators, which often produce motion speeds that are unstable and unpredictable. Key features: - Predicts the Physical Frames Per Second (PhyFPS) of videos. - Corrects the visual motion to align with the true physical frame rate. - Reveals that video generators suffer from chronometric hallucination, producing motion speeds that are unstable and uncontrollable. - Meets the needs of users who want videos that feel more natural and aligned with the physical world. Tech stack: The project is built using Python, with a focus on machine learning and video processing techniques. Why it matters: The Pulse of Motion project is significant because it provides a more accurate and natural way to measure and visualize video motion, addressing the limitations of current video generation methods.