Projects

Overview: TensorFlow Quantum (TFQ) is an open-source Python framework designed for hybrid quantum-classical machine learning, focusing on modeling quantum data. It integrates with Cirq for quantum circuit definitions and qsim for quantum circuit simulations, providing a seamless workflow for leveraging Google's quantum computing offerings. Key features: - Integrates with Cirq for writing quantum circuit definitions - Integrates with qsim for running quantum circuit simulations - Uses Keras for high-level abstractions for quantum machine learning constructs - Provides an extensible system for automatic differentiation of quantum circuits - Offers methods for computing gradients, including parameter shift and adjoint methods - Implements operations as C++ TensorFlow Ops, ensuring they are first-class citizens in the TF compute graph - Harnesses TensorFlow's computational machinery for exceptional performance and scalability Tech stack: Python Why it matters: TensorFlow Quantum empowers quantum algorithms and machine learning researchers to tackle questions that can only be answered through the fast simulation of many millions of moderately-sized circuits, making it essential for ground-breaking research in quantum machine learning. Repository: tensorflow/quantum Primary language: Python GitHub stars: 2198 GitHub description: An open-source Python framework for hybrid quantum-classical machine learning.

Overview: This repository contains Jupyter notebooks designed for learning and practicing Keras. It includes examples on image augmentation, functional API, VGG16 model, pre-trained models, and more. Key features: - Image augmentation techniques - Keras functional API - VGG16 model from scratch - Pre-trained models for classification - Small dataset training - Pre-trained models for classification - Visualizing what convolutional networks learn - Autoencoders - Sequence-to-sequence learning - RNN introduction - LSTM return sequences and states - Learning English alphabet sequence with LSTM - RNN introduction - LSTM for learning English alphabet sequence Tech stack: Python 3.6, Keras 2.1.1, TensorFlow Why it matters: This repository provides practical examples for beginners and intermediate learners of deep learning, helping them better understand and apply Keras in various tasks, including image classification and sequence-to-sequence learning.

Overview: ISAT_with_segment_anything is an interactive semi-automatic image annotation tool based on the Segment Anything model, supporting SAM, SAM2, SAM3, sam-hq, and MobileSAM. It provides visual and text prompts for better annotation efficiency. Key features: - Supports SAM, SAM2, SAM3, sam-hq, and MobileSAM models - Provides visual prompt functionality - Offers text prompt functionality for multiple categories - Supports plugin system for extending functionality with minimal code Tech stack: Python Why it matters: This tool simplifies the process of image segmentation annotation, making it more accessible and efficient for developers and researchers working in the field of computer vision.

Overview: PanWatch is an AI-powered stock monitoring tool designed for A-shares, Hong Kong, and US markets. It integrates with TradingAgents to provide portfolio insights, real-time alerts, and automated reports. Key features: - Portfolio management - AI-scored ideas - Technical indicators - Price alerts - Portfolio manager decision memo Tech stack: Python Why it matters: PanWatch offers a comprehensive solution for investors looking to monitor their portfolios and receive actionable insights from AI-driven analysis, making it easier to make informed decisions in real-time.

### Overview: Harvey is an autonomous AI sales agent powered by Claude Code or Codex. It finds prospects, writes cold emails, sends campaigns, handles replies, and books meetings without requiring API keys. Running on your Claude or Codex subscription, it uses your existing subscription plan to function, eliminating the need for API key costs. ### Key features: - **Prospecting:** Harvey searches for businesses worth selling to, profiles them, and sends cold emails. - **Writing Emails:** It generates personalized cold emails that reference specific details from the prospects. - **Campaigns:** Harvey sends targeted campaigns and handles replies to engage with prospects. - **Meeting Booking:** It handles scheduling meetings and follows up with the responses. ### Tech stack: Python Harvey is built using Python, leveraging the capabilities of Claude Code or Codex through the CLI. ### Why it matters: Harvey's subscription-based model ensures that it doesn't incur per-token costs, making it more cost-effective and practical for businesses looking to automate their sales processes. It also allows for personalized interactions without the need for additional setup or API key management. ### Repository: ethanplusai/harvey The repository is hosted on GitHub, with 1,910 stars, indicating significant interest and adoption in the project. It is maintained by the developer ethanplusai, focusing on automating sales processes through AI. ### Project structure: The project is organized into a modular structure, with clear documentation and examples provided in the README. The codebase is well-documented and easy to understand, making it suitable for developers

### Overview: Reformer, the efficient Transformer, is a Pytorch implementation that offers a memory-efficient approach to natural language processing tasks. It includes LSH attention, reversible network, and chunking, making it suitable for auto-regressive tasks like language modeling. ### Key features: - **Reformer Language Model**: A simple Reformer language model that can fit into 5-8k tokens. - **LSH Attention**: A memory-efficient attention mechanism that reduces memory usage. - **Reversible Network**: A reversible network that enhances the speed of the model without compromising accuracy. - **Chunking**: The model processes sequences in chunks to manage memory efficiently. ### Tech stack: Pytorch ### Why it matters: Reformer is particularly efficient for tasks that require high memory usage, such as language modeling, making it a valuable tool in the field of natural language processing. Its ability to handle large sequences with minimal memory footprint makes it a standout choice for applications where computational resources are a limiting factor.

Overview: MoBA is a novel approach to mixture of block attention for long-context language models, designed to efficiently process large sequences without the drawbacks of traditional attention mechanisms. Key features: - Trainable Block Sparse Attention - Parameter-less Gating Mechanism - Seamless Transition Between Full and Sparse Attention Tech stack: Python Why it matters: MoBA offers a flexible and efficient solution for long-context tasks, improving the performance and scalability of large language models without the need for additional training or modification of pretrained models. Repository: MoonshotAI/MoBA GitHub stars: 2194 GitHub description: MoBA: Mixture of Block Attention for Long-Context LLMs

### Overview: AI Agents hallucinate. Fix it faster. The open-source platform for evaluating, observing, and improving LLM and AI agent applications. ### Key features: - Evaluations - Tracing - Simulations - Datasets - Gateway - Guardrails ### Tech stack: Python ### Why it matters: Self-hostable, Apache 2.0 license, and optimized for continuous improvement and deployment.

### Overview: TFX is an end-to-end platform for deploying production ML pipelines, providing a configuration framework for TensorFlow components and enabling advanced functionality like experiment tracking and warmstarting/resuming models from previous runs. ### Key features: - Configuration framework for TensorFlow pipelines - Integration with Apache Airflow and Kubeflow Pipelines - ML Metadata backend for component runs, input/output artifacts, and runtime configuration - Supports advanced features like experiment tracking and warmstarting/resuming models from previous runs ### Tech stack: Python ### Why it matters: TFX is an essential tool for anyone working with TensorFlow and machine learning pipelines, providing a robust and scalable solution for deploying production ML models.

Overview: node-llama-cpp is a project that allows developers to run AI models locally on their machines using node.js bindings for llama.cpp. It provides pre-built binaries for macOS, Linux, and Windows, and supports GPU, Metal, and Vulkan. Key features: - Run LLMs locally on your machine - Metal, CUDA, and Vulkan support - Pre-built binaries provided, with a fallback to building from source without node-gyp or Python - Automatically adapts to your hardware, no need to configure anything - Complete suite of everything needed for using LLMs in projects - Use CLI to chat with a model without installing - Up-to-date with the latest llama.cpp release - Enforce model output in JSON or specific JSON schema - Provide model with functions it can call on demand - Embedding and reranking support - Structured decisions support with any model - Safe against special token injection attacks - Great developer experience with TypeScript support and complete documentation - Much more Tech stack: TypeScript Why it matters: node-llama-cpp simplifies the process of integrating AI models into projects by providing a local, easily installable solution. It supports multiple hardware platforms and ensures model output is in a structured format, making it easier for developers to integrate AI into their applications.

Overview: This repository contains a package designed for machine learning inference on FPGAs using High-Level Synthesis (HLS). It allows for the creation of firmware implementations of machine learning algorithms on FPGAs. Key features: - HLS-based machine learning inference - Support for various scientific and industrial applications - Example use cases including control systems for quantum computing, feedback loops in nuclear fusion, and biomedical signal processing Tech stack: Python Why it matters: This package is designed for ultra-low-latency inference on FPGAs, making it suitable for applications requiring high performance and low latency, such as in the field of quantum computing and environmental monitoring. Repository: fastmachinelearning/hls4ml GitHub stars: 2191 GitHub description: Machine learning on FPGAs using HLS [](https://doi.org/10.5281/zenodo.19134628) [](https://opensource.org/licenses/Apache-2.0) [](https://fastmachinelearning.org/hls4ml) [](https

### Overview: Handwritten Text Recognition (HTR) system implemented with TensorFlow, trained on the IAM off-line HTR dataset. The model recognizes text on single words or text lines (multiple words) and achieves 3/4 correct recognition and a character error rate of around 10%. ### Key features: - Recognizes text on single words or text lines (multiple words). - Outputs recognized text with high accuracy. - Recognizes words on line level, improving recognition on text lines. - Supports multiple-word input images. - Compatible with TensorFlow 2.x. ### Tech stack: Python ### Why it matters: This system is useful for applications that require accurate text recognition from handwritten or printed text, such as document scanning, OCR, and other recognition tasks. The high accuracy and ability to handle multiple words make it suitable for various practical applications.

Overview: pgvecto.rs is a Postgres extension that provides scalable, low-latency, and hybrid-enabled vector search functions, revolutionizing vector search capabilities in databases. Key features: - Supports up to 65,535 dimensions for vector search. - Handles filtering with VBASE method for vector search and relational query. - Utilizes SIMD instructions for dynamic performance optimization. - Introduces additional data types like binary vectors, FP16, and INT8. - Supports indexing and storage of indexes separately from PostgreSQL. - Provides Write-Ahead Logging (WAL) support for data and index. Tech stack: Rust, pgrx. Why it matters: pgvecto.rs offers a significant improvement in vector search capabilities by providing better performance and stability compared to existing solutions like pgvector. It is designed to be easily integrated into existing PostgreSQL environments, making it a valuable tool for developers looking to enhance their database search functionalities.

### Overview: llama.vim is a Vim plugin designed for LLM-assisted code/text completion, providing auto-suggest, instruction-based editing, and enhanced context management. ### Key features: - **Auto-suggest on cursor movement in `Insert` mode** - **Accept a suggestion with `Tab`** - **Accept the first line of a suggestion with `Shift+Tab`** - **Instruction-based editing with `<leader>lli`** - **Control max text generation time** - **Configure scope of context around the cursor** - **Ring context with chunks from open and edited files and yanked text** - **Supports very large contexts even on low-end hardware via smart context reuse** ### Tech stack: Vim Script ### Why it matters: llama.vim enhances your coding experience by providing intelligent suggestions and context-aware editing, making your development more efficient and productive.

Overview: Pretzel is a modern, open-source alternative to Jupyter Notebooks designed to enhance the capabilities of Jupyter. Key features: - AI code generation and editing - Inline tab completion - Sidebar chat and error fixing for Jupyter cells - Native AI code generation and understanding features similar to Cursor - Frictionless realtime collaboration - SQL support (both in code cells and as a standalone SQL IDE) - Visual analysis builder - VSCode-like code-writing experience using Monaco Editor - 1-click dashboard creation and sharing from Jupyter notebooks Tech stack: TypeScript Why it matters: Pretzel aims to improve Jupyter's capabilities by adding AI-driven features, making it easier and more efficient for developers to work with Jupyter notebooks.

Overview: This repository contains a PyTorch implementation of adversarial attacks, providing a user-friendly interface for generating adversarial examples. Key features: - Adversarial attacks for various PyTorch models - Easy-to-use interface for implementing adversarial attacks - Supports different types of perturbations and attack methods Tech stack: Python Why it matters: This package simplifies the process of applying adversarial attacks to PyTorch models, making it easier for developers to experiment with and improve their models' robustness.

Overview: EasyNLP is a comprehensive and user-friendly NLP toolkit built on PyTorch, designed for easy development and application in Alibaba Group. It supports knowledge distillation, few-shot learning, and various multi-modality pre-trained models, making it highly adaptable and efficient for real-world NLP tasks. Key features: - Easy-to-use commands for cutting-edge models - Customizable modules like AppZoo and ModelZoo - TorchAccelerator for distributed training speed - Support for Huggingface/Transformers and EasyTransfer ModelZoo - Knowledge-injected pre-trained models - Few-shot learning for model fine-tuning - Knowledge distillation for model size reduction - Multi-modality pre-trained models for vision-language tasks Tech stack: PyTorch, distributed training, knowledge distillation, few-shot learning, multi-modality pre-trained models Why it matters: EasyNLP simplifies the development and deployment of NLP applications within Alibaba Group, enabling efficient and effective use of large pre-trained models and knowledge-injected pre-training techniques.

### Overview: This repository is a curated collection of papers and blogs on long-context language modeling, covering surveys, efficient attention mechanisms, KV-cache optimization, recurrent transformers, state-space models, position encoding, long-context training, long-term memory, retrieval-augmented generation, in-context learning, context and model compression, long reasoning, long video & image, long-horizon agents, long-text generation, inference acceleration, benchmarks & evaluation, and technical reports. ### Key Features: - Surveys on long-context language modeling - Efficient attention mechanisms - KV-cache optimization - Recurrent transformers and state-space models - Position encoding and length extrapolation - Long-context training - Long-term memory - Retrieval-augmented generation - In-context learning - Context and model compression - Long reasoning (long CoT) - Long video & image - Long-horizon agents - Long-text generation - Inference acceleration - Benchmarks & evaluation ### Tech Stack: - Python, JavaScript, Markdown, LaTeX, HTML, CSS ### Why It Matters: This repository serves as a valuable resource for researchers and practitioners interested in the field of long-context language modeling, providing comprehensive insights and references to stay updated with the latest advancements in the area.

### Overview: Gaussian Splatting SLAM is a dense simultaneous localization and mapping (SLAM) system presented at CVPR 2024. The method demonstrates the first monocular SLAM system based on 3D Gaussian splatting, which also supports stereo/RGB-D inputs. It is highlighted in CVPR'24 as a CVPR 2024 Highlight and Best Demo Award. ### Key Features: - **Monocular SLAM System**: Implements dense SLAM using monocular vision, showcasing the first monocular SLAM solely based on 3D Gaussian splatting. - **Stereo/RGB-D Support**: Supports stereo/RGB-D inputs, providing a more comprehensive SLAM solution. - **Dense SLAM**: Achieves dense SLAM, enabling accurate mapping and localization in dense environments. - **3D Gaussian Splatting**: Utilizes 3D Gaussian splatting for accurate 3D reconstruction, providing high-quality mapping. ### Tech Stack: - **Programming Language**: Python - **Dependencies**: PyTorch, CUDAToolkit, OpenCV, NumPy, Matplotlib, and other standard Python libraries. ### Why It Matters: - **Innovative SLAM Method**: The method represents a significant advancement in monocular SLAM, offering a unique approach to dense SLAM. - **Highlight at CVPR 2024**: Demonstrates the first monocular SLAM system based on 3D Gaussian splatting, showcasing its potential in dense environments. - **Versatility**: Supports both

### Overview: Finding a flight or hotel shouldn't mean checking 47 tabs or spending hours searching. Our platform provides a single, reliable price for every flight, eliminating the need for tracking and price fluctuations. ### Key Features: - **Agent-Native Flight Search and Booking**: Our server-side engine scans the entire world for the cheapest price, ensuring every flight is always the best deal. - **Hotel Booking Integration**: Book hotels directly from your AI agent, with real PNRs and guaranteed availability. - **Worldwide Coverage**: Every major airline and major booking sites are included, ensuring you can find the best deals on every flight. ### Tech Stack: Python ### Why It Matters: Every airline in the world, major booking sites, and OTAs are included, ensuring you can find the best deals on every flight. Our server-side engine scans the entire world for the cheapest price, eliminating the need for tracking and price fluctuations. The same flight costs $20-$50 less because you ski, making it easier and more convenient for your AI agent to search and book flights.