Projects

Overview: Surround is a lightweight framework designed for building AI-driven microservices in Python, focusing on ease of use and problem-solving for data scientists. Key features: - Easy to use - Flexible architecture - Supports multiple models and glue code integration - Allows for the evolution of machine learning pipelines without re-engineering Tech stack: Python Why it matters: Surround simplifies the process of integrating machine learning pipelines into microservices, making it accessible to data scientists without the need for extensive coding. Repository: a2i2/surround Primary language: Python GitHub stars: 23 GitHub description: Surround is a lightweight framework for serving machine learning pipelines in Python. It is designed to be flexible, easy to use, and focused on problem-solving for data scientists. Repository: a2i2/surround Primary language: Python GitHub stars: 23 GitHub description: Surround is a lightweight framework for serving machine learning pipelines in Python. It is designed to be flexible, easy to use, and focused on problem-solving for data scientists.

### Overview: MinimalGPT is a concise, adaptable, and streamlined code framework designed specifically for the construction, training, inference, and fine-tuning of the GPT model. This framework is implemented using Keras and TensorFlow, ensuring compatibility and coherence within the broader deep learning ecosystem. ### Key Features: - **MinimalGPT Framework**: A comprehensive, adaptable, and streamlined code framework for GPT model construction, training, inference, and fine-tuning. - **Python Implementation**: Utilizes Keras and TensorFlow, ensuring compatibility and coherence within the deep learning ecosystem. - **Command-Line Interface**: Simplifies the process of model creation, training, saving, loading, fine-tuning, and inference, all through a single command line execution. ### Tech Stack: - Python 3.10 - Keras - TensorFlow ### Why It Matters: MinimalGPT provides a straightforward, efficient, and versatile solution for GPT model development, making it an ideal choice for developers looking to streamline their deep learning projects.

### Overview: Build a Large Language Model (LLM) from Scratch ### Key Features: - Code and resources for building a large language model from scratch. - Guided by Sebastian Raschka's book "Build a Large Language Model (from Scratch)." - Demystifies the process of creating, training, and fine-tuning LLMs. - Provides a hands-on approach to understanding the underlying mechanics of LLMs. ### Tech Stack: Jupyter Notebook ### Why it Matters: Gain a comprehensive understanding of LLMs and develop your own functional model.

### Overview: This repository contains an awesome directory of AI tools and applications, curated from various sources. It serves as a searchable interface for those interested in enhancing their AI journey. ### Key features: - **AI Japanese Tutor**: A free AI voice chat bot for practising Japanese conversation and speaking skills. - **NeatFlow AI**: A convenient platform to use ChatGPT, Claude & Gemini. - **Character.AI**: Create and chat with AI characters and personalities. - **Replika**: An AI companion focused on emotional support and conversation. - **AI Image Generator**: A free AI-powered text-to-image and image-to-image art generator. - **Lexica**: An AI art search engine and generation platform. - **Neural.love**: AI art generator with historical photo restoration. - **NightCafe**: AI art creation with multiple style options and community features. - **Stable Horde**: Distributed volunteer-powered image generation. - **AI Logo Generator**: A free online AI logo creator for professional company, business, and brand logos. - **AltTextLab**: AI-powered tool for automatic alt text generation for images. ### Tech stack: GitHub ### Why it matters: This repository serves as a valuable resource for developers and enthusiasts looking to explore and utilize a wide range of AI tools and applications. It provides a searchable interface for those interested in enhancing their AI journey, making it an essential tool for anyone looking to supercharge their AI capabilities.

### Overview: This project leverages OpenAI and ElasticSearch to provide a Q&A-like experience on long-form text, enhancing the retrieval and generation capabilities through a Retrieval-Augmented Generation (RAG) approach. ### Key features: - **ElasticSearch as Vector Database**: Stores and retrieves indexed data for efficient querying. - **OpenAI Embeddings**: Used to generate answers by leveraging embeddings derived from the indexed data. - **Python Backend**: Implements the application using Python, virtual environment, and required dependencies. ### Tech stack: Python ### Why it matters: This project offers a unique Q&A experience by combining the power of OpenAI's AI capabilities with the advanced indexing and retrieval capabilities of ElasticSearch, making it suitable for applications requiring both high relevance and generation quality in response to user queries.

### Overview: This repository contains optimized implementations of the Jaccard Index, focusing on improving performance for large-scale vector search. It leverages techniques such as lookup tables, Harley-Seal CSA, and precomputed Jaccard JUTs to enhance the efficiency of operations involving bit-level representations. ### Key Features: - **Lookup Tables for Population Counts**: Speeds up the calculation of population counts by precomputing and storing results. - **Harley-Seal CSA for Population Counts**: Utilizes Odd-Majority Carry-Save Adders to efficiently compute population counts. - **Loop Unrolling and Inlining**: Reduces the number of instructions and speeds up the execution. - **Precomputed Jaccard JUTs**: Provides precomputed results for quick lookups. - **Floating-Point Operations**: Replaces integer operations for better performance. ### Tech Stack: - Python: Implementations using NumPy, Numba, and C++ using Cppyy JIT compiler. ### Why It Matters: - The repository benchmarks and optimizes the performance of Jaccard Index calculations, particularly for large-scale vector search applications. By leveraging precomputed results and optimized kernels, it significantly improves the efficiency of operations involving bit-level representations, making it highly relevant for large-scale quantized vector search deployments.

Overview: 小红书图文笔记 Agent Skill，支持图片理解、主动提问、智能选题、实时热梗搜索、标题正文、标签评论钩子和风险检查。 Key features: - 图片分析 - 主动提问 - 智能选题 - 实时热梗搜索 - 标题正文 - 标签评论钩子 - 风险检查 Tech stack: Python Why it matters: 该技能为小红书图文内容提供全面的分析和生成功能，帮助用户从选题到复盘的全流程优化笔记质量，提升内容的吸引力和可分享性。 Repository: azlll/xhs-notes-skill Primary language: Python GitHub stars: 22

Overview: This project implements a multi-agent system for zero-shot visual question answering (VQA) using foundation models. The system explores the capabilities of foundation models in zero-shot scenarios and proposes an adaptive multi-agent approach to improve performance. Key features: - Multi-Agent VQA: An adaptive multi-agent system for zero-shot VQA. - Foundation Models: Foundation models are used as tools to overcome limitations in object detection and counting. - Zero-Shot Scenario: The system is tested in zero-shot scenarios without fine-tuning on specific VQA datasets. Tech stack: Python Why it matters: This work explores the zero-shot capabilities of foundation models in VQA tasks, providing a practical and robust solution for visual question answering in the open world. Repository: bowen-upenn/Multi-Agent-VQA GitHub stars: 23 GitHub description: This is the official implementation of the paper ["Multi-Agent VQA: Exploring Multi-Agent Foundation Models in Zero-Shot Visual Question Answering"](https://arxiv.org/abs/2403.14783) in Pytorch. Key idea: The paper proposes an adaptive multi-agent system to overcome limitations in foundation models, using specialized agents as tools. The system is tested in zero-shot scenarios, making it more practical and robust in the open world. Disclaimer: This repository is under development, and currently only supports GPT-4V and Gemini Pro Vision as large vision-language models. Instructions are provided for other models and functionalities, but they are either incomplete or not thoroughly tested yet. TODOs: 1. Int

### Overview: This project aims to simplify and summarize legal documents using Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG). It leverages information retrieval techniques to extract key insights from complex agreements, contracts, and other legal texts, making them more accessible to non-lawyers. ### Key Features: - **Retrieval-Augmented Generation (RAG)**: Enhances the summarization process by first searching for relevant content and then reconstructing it using an LLM. - **Document Processing**: Extracts information using OCR, transformers, and keyword tagging. - **Document Retrieval**: Uses BM25 ranking or Semantic Search to fetch relevant parts of legal documents. - **Context Augmentation**: The retrieved text is passed to an LLM to generate a structured and readable summary. ### Tech Stack: - Python: The primary programming language used for the project. - Streamlit: A Python library for building web applications. ### Why It Matters: - Simplifies understanding complex legal documents for non-lawyers, reducing the risk of misinterpretation and serious consequences. - Improves accessibility of legal content, encouraging individuals to handle contracts themselves. - Enhances the efficiency and accuracy of legal document processing, making it easier and faster to extract key information.

Overview: ragctl is a powerful CLI tool designed for managing, testing, and optimizing Retrieval-Augmented Generation (RAG) pipelines. It streamlines the process of document ingestion, OCR, and chunking, allowing developers to focus on building their RAG applications. Key features: - Universal document loading (PDF, DOCX, images, HTML, Markdown, etc.) - Advanced OCR with fallback options (EasyOCR → PaddleOCR → pytesseract) - Intelligent semantic chunking using LangChain - Production-ready batch processing with auto-retry - Multiple export formats (JSON, JSONL, CSV) - Direct ingestion into Qdrant vector store Tech stack: Python Why it matters: ragctl simplifies the complex process of preparing documents for RAG systems, making it easier for developers to focus on the core functionality of their applications. Repository: datallmhub/ragctl GitHub stars: 23 GitHub description: A powerful CLI tool to manage, test, and optimize RAG pipelines. Streamline your Retrieval-Augmented Generation workflows from terminal.

Overview: This project is an official implementation of a paper that introduces a comprehensive and extensible "Video Toolkit" for enhancing spatiotemporal reasoning capabilities in MLLM, specifically for streamlining video question answering tasks. Key features: - Video Toolkit - Spatiotemporal Reasoning Framework (STAR) - Tools for temporal grounding, temporal QA, temporal referencing, and image QA Tech stack: Python Why it matters: The project showcases advancements in using lightweight tools to enhance language models, specifically improving spatiotemporal reasoning in video question answering tasks, resulting in significant performance gains. Repository: fansunqi/VideoTool GitHub stars: 23 GitHub description: Official Repository for NeurIPS'25 Paper "Tool-Augmented Spatiotemporal Reasoning for Streamlining Video Question Answering Task"

### Overview: This repository provides core code for managing large volumes of video footage, enabling content understanding, automatic tagging, and vector database storage. It integrates multimodal models and large language models for accurate descriptions and semantic search. A web interface allows visualization. ### Key features: - **Video Scene Understanding (Local Deployment)**: Uses *Google Gemma3* to analyze the entire video's dynamic information, including people, objects, and scenes. - **Audio Analysis (Local Deployment)**: Employ *sensevoice* for audio analysis and extracting dialogue, providing textual information for content summarization and tag generation. - **Tag and Video Description Generation (via API)**: Based on multimodal information, a large language model generates tags, summaries, or descriptions. - **Database Storage and Retrieval (Local Deployment)**: Uses structured database (SQLite) and vector database (ChromaDB) to store video analysis results for efficient retrieval. - **Natural Language Video Search (Local Deployment)**: Uses query.py to implement natural language queries, finding the most matching videos through vector similarity search. - **Web Interface Management and Search (Local Deployment)**: Provides a Flask-based web interface supporting video browsing, searching, previewing, and exporting. - **Search and Management (Local Deployment)**: Combines with Adobe Bridge or similar tools to search for video clips by keywords. - **Creative Content Production**: Generates script outlines for platforms like Rednote or Tiktok based on video descriptions, tags, and extracted dialogue. ### Tech stack: Python ### Why it matters: This repository is an intelligent tool that addresses two major

Overview: This repository contains the code for the ICCV 2023 paper "XVO: Generalized Visual Odometry via Cross-Modal Self-Training". It includes the implementation of a semi-supervised learning method for training generalized monocular Visual Odometry models. Key features: - XVO: Generalized Visual Odometry via Cross-Modal Self-Training - Semi-supervised training for robust learning - Multi-modal supervision (segmentation, flow, depth, and audio) Tech stack: Python Why it matters: This method demonstrates the benefits of semi-supervised learning and multi-modal supervision in improving the robustness and generalizability of visual odometry models across diverse datasets and settings. Repository: h2xlab/XVO GitHub stars: 23 GitHub description: The repository contains the code for the ICCV 2023 paper "XVO: Generalized Visual Odometry via Cross-Modal Self-Training".

Overview: AnimAI-Trainer is a project focused on training an LLM to generate cracked Manim animations for mathematical concepts. Key features: - Data processing - Model training Tech stack: Python Why it matters: This project aims to automate the creation of educational visual aids using AI, making complex mathematical concepts more accessible through animated visualizations. Repository: hesamsheikh/AnimAI-Trainer GitHub stars: 23

Overview: R-LoRA is an innovative project that introduces Random Initialization of Multi-Head LoRA for Multi-Task Learning, aiming to improve the performance of LoRA in complex multi-task scenarios. Key features: - Incorporates Multi-Head Randomization to enhance task-specific learning. - Provides reproducible code for multi-task adaptation. - Demonstrated superior performance in capturing task-specific knowledge. Tech stack: Python Why it matters: R-LoRA offers a cost-effective solution for multi-task learning by diversifying head matrices, leading to more efficient and effective fine-tuning of large language models. Repository: jinda-liu/R-LoRA GitHub stars: 23

### Overview: Learn the ins and outs of efficiently serving Large Language Models (LLMs). Dive into optimization techniques, including KV caching and Low Rank Adapters (LoRA), and gain hands-on experience with Predibase’s LoRAX framework inference server. ### Key features: - Auto-Regressive Models - LLM Inference Stack - Low Rank Adapters (LoRA) - Hands-On Experience with Predibase’s LoRAX framework inference server ### Tech stack: Jupyter Notebook ### Why it matters: Gain deep insights into optimizing LLM applications for large-scale deployment. Understand techniques like KV caching and Low Rank Adapters (LoRA) to efficiently serve LLMs to a large number of users. Experience hands-on with Predibase’s LoRAX framework inference server to see optimization techniques in action.

Overview: Smallevals is a lightweight evaluation framework designed for offline retrieval evaluation of RAG systems using tiny 0.6B models. It runs efficiently on CPU, GPU, and MPS, making it ideal for local testing and evaluation. Key features: - Supports evaluation of retrieval and RAG stages. - Runs 100% locally on CPU, GPU, and MPS. - Fast and free evaluation, with results available in seconds on GPU and minutes on CPU. - Evaluates models for retrieval quality, context relevance, and answer correctness. Tech stack: Python. Why it matters: Smallevals simplifies the evaluation process for RAG systems, making it accessible and cost-effective, especially for small QA models. Repository: mburaksayici/smallevals GitHub stars: 23 GitHub description: smallevals — CPU-fast, GPU-blazing fast offline retrieval evaluation for RAG systems with tiny QA models. Logo: smallevals cat logo Primary language: Python Installation: `pip install smallevals` Quick Start: Evaluate retrieval quality by connecting to a Vector DB, generating questions, and visualizing results. For ChromaDB, use `from smallevals import evaluate_retrievals, SmallEvalsVDBConnection`. ChromaDB example: `from sentence_transformers import SentenceTransformer, SmallEvalsVDBConnection, evaluate_retrievals`. Elasticsearch example: `from elasticsearch import Elasticsearch, SmallEvalsVDBConnection, evaluate_retrievals`.

Overview: SpectralGCD is an efficient and effective multimodal approach to Generalized Category Discovery (GCD) that uses CLIP cross-modal image-concept similarities as a unified cross-modal representation. It addresses the limitations of traditional methods by leveraging semantic information and reducing overfitting. Key features: - Spectral Filtering: A technique that automatically retains only relevant concepts from a large task-agnostic dictionary, ensuring semantic quality and well-aligned cross-modal representations. - Cross-modal Representation Learning: A unified cross-modal representation using CLIP to improve performance and reduce computational cost. Tech stack: Python with PyTorch and CUDA 12.4. Why it matters: SpectralGCD delivers state-of-the-art accuracy at a fraction of the computational cost, making it a significant advancement in the field of GCD. Repository: miccunifi/SpectralGCD GitHub stars: 23 GitHub description: [ICLR 2026] - Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery.

Overview: This repository is a comprehensive interview prep curriculum organized by career path, focusing on the job you're actually interviewing for. It includes five routes: Data Structures & Algorithms (DSA), Behavioral & Reverse Interview, System Design & Computer Science Fundamentals, Machine Learning (Classical ML), and LLM Post-Training. Key features: - Suggested order for each career path - A question bank for each route - Open links to specific READMEs for each route - Resume tips PDF for getting ready for the interview Tech stack: Java Why it matters: This repo is designed to save time by providing a focused, career-specific interview prep experience. It's organized to help you focus on the material relevant to your specific job path, ensuring that you're prepared for the right questions and challenges. Repository: mostofashakib/tech-interview-curriculum GitHub stars: 23 GitHub description: Interview prep for LLM post-training, ML, system design, and coding — organized by career path.

### Overview: This project is an enterprise-grade multi-agent AI customer support system that has been designed to save companies £150K+ annually. It includes a complete tutorial series and is built with Python, featuring a multi-agent architecture, RAG + database hybrid, location intelligence, safety guardrails, and production-ready features. ### Key Features: - **Multi-Agent Architecture**: Intelligent routing between specialized agents. - **RAG + Database Hybrid**: Best of both worlds for knowledge management. - **Location Intelligence**: Smart defaults based on customer location. - **Safety Guardrails**: PII protection, prompt injection defense, content filtering. - **LLM-as-a-Judge Evaluation**: Automated quality assessment with detailed reports. - **Production Ready**: 15 test scenarios, deployment docs, error handling. - **Proven ROI**: Companies save £150K+ annually vs traditional support teams. ### Tech Stack: Python ### Why it Matters: This system is designed to be enterprise-grade, ensuring seamless integration with existing infrastructure. The complete tutorials included make it an ideal learning experience, providing a comprehensive guide to building and deploying an AI customer support system.