Projects

Overview: Praetor is a lightweight finetuning data and prompt management tool designed for LLMs, offering a user-friendly interface for exploring, editing, and exporting datasets. Key features: - Dataset exploration and management - Prompt style viewer and editor - JSON export and import capabilities - Minimal setup requirements Tech stack: Python Why it matters: Praetor simplifies the process of managing and exploring datasets for LLM finetuning, making it easier to work with large JSON files and control F searches. Repository: goodreasonai/praetor-data Primary language: Python GitHub stars: 67 GitHub description: Praetor is a dataset exploration tool. It is made for LLM finetuning data and prompts. README excerpt: # Praetor Praetor is a dataset exploration tool. It is made for LLM finetuning data and prompts. ## Screenshots  ---  ## Why I got sick of scanning through large json files to try and get a feeling for a dataset. Then when I wanted to remove some data, I was using control F - this is obviously dumb. Praetor is a flask app that lets you mix, edit, view, and export datasets for LLMs. To keep it simple, the frontend is very straightforward. It uses an sqlite database and jinja templating under the hood. You

### Overview: Condo GPT is an intelligent assistant designed to query and analyze condominium data in Miami, providing insights and comparisons to help real estate agents and investors make informed decisions. ### Key Features: - **Natural Language Interface:** Interprets user questions and provides answers. - **Integration with Google Maps API:** Facilitates location-based queries. - **Dynamic SQL Generation:** Enables complex database queries. - **Interactive Maps and Charts:** Visualizes data for better understanding. - **PDF Report Generation:** Offers detailed analysis in a report format. ### Tech Stack: Python, Flask, PostgreSQL, LangChain, LangGraph, OpenAI's GPT models, Google Maps API, Chart.js, ReportLab. ### Why it Matters: This tool simplifies the process of querying and analyzing condominium data, making it accessible to non-technical professionals. By automating the analysis, it can save real estate agents and investors thousands of dollars in salary, allowing them to focus on their core business.

### Overview: ITI-GEN aims to generate images that represent diverse attributes uniformly across categories, using only marginal distribution datasets. ### Key features: - Only need datasets that capture the marginal distribution of individual attributes. - Learned token embeddings are generalizable across different generative models. ### Tech stack: Python ### Why it matters: - Generates images that cover a wide range of attributes uniformly, without the need for joint distribution datasets. - Improves the diversity and inclusivity of image generation.

Overview: This repository provides an efficient implementation of prefix beam search for the RNN-Tranducer loss function, optimized for real-time applications. It is designed to achieve near-instantaneous processing times, making it suitable for production-level error rates. Key features: - Optimized prefix beam search implementation for RNN-Tranducer loss function. - Fast processing time (~100 ms for a 5-second speech segment with a beam size of 10). - Supports real-time applications with minimal latency. Tech stack: C++/CUDA Why it matters: This implementation is crucial for real-time speech recognition systems, enabling faster and more efficient processing of speech data compared to traditional methods. Repository: iamjanvijay/rnnt_decoder_cuda GitHub stars: 67

### Overview: These tutorials provide deep dives into AI agents, LLM frameworks, coding tools, and open-source developer platforms, explaining how complex systems work under the hood with architecture diagrams, real code walkthroughs, and production-grade patterns. ### Key Features: - Architecture diagrams - Annotated code examples - Real code walkthroughs - Production-grade patterns - Consistent 8-chapter structure - Mermaid architecture diagrams - Annotated code examples - Summary tables ### Tech Stack: Python ### Why It Matters: These tutorials help developers understand complex systems and their underlying architecture, providing valuable insights into how to design, extend, and optimize systems effectively.

### Overview: ZeusHammer is an AI super agent that combines the power of three open-source giants—ClaudeCode, Hermes, and OpenClaw—to deliver unparalleled performance, privacy, and local processing capabilities. It offers a 100% local processing model, a 3-tier memory system, and a whisperable voice interaction, making it the ultimate AI solution for developers seeking cost savings, speed, and privacy. ### Key Features: - **Local Brain**: Handles 80% of tasks locally, reducing cloud costs and speeds. - **Voice Interaction**: Seamless voice-based communication with whisperable technology. - **Three-Tier Memory**: Short-term, long-term, and working memory for persistent data storage. - **Meditation Mode**: Auto-analysis and skill upgrades on idle time. - **Deep Reflection**: Causal chain analysis and insight generation. - **Workflow Engine**: Complete orchestration, skill matching, and skill learning. - **Circuit Breaker**: Comprehensive security system. - **OSV Scanning**: Command/file malware detection. - **China LLM Native**: Built-in DeepSeek/Qwen/GLM/Moonshot/ERNIE/Doubao/MiniMax. - **Tool Detector**: Auto-detects 80+ system CLI tools. - **Cost Tracking**: Built-in LLM cost tracking. - **Smart Model Router**: Intelligent routing, fallback chains, and cost optimization. - **WebRTC Real-time Audio**: Real-time voice streaming. - **Collaboration Rooms**: Multi-user collaboration. - **Apple Ecosystem**: Integration with Apple Notes

Overview: PhyT2V is an innovative text-to-video generation technique that leverages large language models to refine prompts, enabling more accurate and physically realistic video outputs. Key features: - Iterative self-refinement using LLMs - Adapts to out-of-distribution domains - Improves adherence to real-world physical rules by 2.3x - Enhances performance compared to T2V prompt enhancers by 35% Tech stack: Python Why it matters: PhyT2V addresses the limitations of current T2V models by incorporating LLM-guided iterative self-refinement, making it a significant advancement in the field of physics-grounded text-to-video generation. Repository: pittisl/PhyT2V Primary language: Python GitHub stars: 67

Overview: This iOS Swift project converts the time heard in the UK to the British way of saying it, providing a British accent and customizable design options. Key features: - Converts time heard in the UK to the British way of saying it. - Allows customization of the clock design (Classic, Art Nouveau, or Drawing Style). - Provides a British accent for spoken time. - Offers a speech recognition feature for pronunciation training. - Displays minute/hour indicators or limited hour as per preference. - Provides a Today Widget for quick time display. Tech stack: SwiftUI, SwiftClockUI, The Composable Architecture (TCA), SwiftPastTen, SwiftToTen, SwiftTTSCombine, SwiftSpeechCombine, SnapshotTesting. Why it matters: This app helps users who are accustomed to the British way of telling time in the UK, providing a British accent and customizable design options, making it easier for them to confirm the time they hear.

### Overview: This project introduces a novel phasic content fusing few-shot diffusion model that addresses the limitations of few-shot model adaptation by incorporating a phasic training strategy and a directional distribution consistency loss. The model is designed to improve the capture of content, style, and local details in few-shot generative tasks, demonstrating superior performance compared to state-of-the-art methods. ### Key features: - Phasic content fusing few-shot diffusion model - Directional distribution consistency loss - Phasic training strategy for different learning objectives at distinct stages - Enhanced structure consistency during domain adaptation ### Tech stack: Python ### Why it matters: This work addresses the challenge of training generative models with limited data by proposing a novel approach that improves the quality of generated content, style, and local details, making it particularly useful in few-shot generative model adaptation tasks. The superior performance demonstrated in the provided results suggests significant potential for practical applications in various domains that require high-quality generative models.

Overview: Agenite is a TypeScript framework designed for building powerful AI agents, providing a modular, type-safe, and flexible architecture that supports various providers like OpenAI, Anthropic, AWS Bedrock, and Ollama. Key features: - Type safety and developer experience - Tool integration and first-class support for function calling - Provider agnostic interface - Advanced architecture with bidirectional flow, step-based execution, and state management - Model context protocol (MCP) for connecting LLMs to data sources Tech stack: TypeScript Why it matters: Agenite simplifies the process of building AI agents by providing a robust, type-safe, and modular framework that supports multiple providers and advanced features like MCP, making it suitable for developers looking to create sophisticated AI solutions with ease and efficiency. Repository: subeshb1/agenite GitHub stars: 67 GitHub description: 🤖 Build powerful AI agents with TypeScript. Agenite makes it easy to create, compose, and control AI agents with first-class support for tools, streaming, and multi-agent architectures. Switch seamlessly between providers like OpenAI, Anthropic, AWS Bedrock, and Ollama.

Overview: SUTD-TrafficQA is a challenging Video Question Answering (VQA) benchmark based on real-world traffic scenes, designed to evaluate the ability of video reasoning over traffic events. Key features: - Real-world traffic scenes - Efficient network for video reasoning - Publicly released dataset Tech stack: JavaScript Why it matters: SUTD-TrafficQA provides a benchmark for evaluating video reasoning capabilities in traffic events, contributing to advancements in autonomous driving and intelligent transportation systems.

### Overview: ToStore is a fast distributed AI vector database and persistent local storage engine designed for modern AGI and edge intelligence scenarios. It offers high-performance key-value storage, SQL, NoSQL, offline cache, and encrypted data support. Built on a self-routing node architecture, ToStore ensures high autonomy and elastic horizontal scalability, decoupling performance from data scale. It seamlessly supports declarative schema changes, key rotation, and data refactoring, ensuring continuous evolution and iteration without service interruptions. ### Key Features: - **Key-Value Storage (KV) Mode**: Supports SQL, NoSQL, offline cache, and encrypted data. - **Structured Table Mode**: Offers advanced querying capabilities, aggregation, statistics, complex logic, reactive queries, streaming queries, advanced key-value operations, bulk operations, recovery durability, vector fields, hybrid retrieval, table-level TTL, efficient pagination, memory probe and sync retrieval (peek), manual query result caching, atomic expressions, and transactions. - **Distributed Architecture**: Built on a self-routing node architecture, ensuring high autonomy and elastic scalability. - **Cascading Foreign Keys**: Supports cascading foreign keys and efficient data management. - **Mobile/Desktop Integration**: Seamless integration for mobile and desktop environments. - **Server/Agent Integration**: Underpins autonomous evolution and continuous iteration for agents and automated O&M. - **Primary Key Algorithms**: Supports various primary key algorithms for efficient data retrieval. ### Tech Stack: Dart ### Why It Matters: ToStore is crucial for modern AI and edge intelligence applications, offering a unified data engine that supports relational structured data, ensuring

Overview: This repository provides a comprehensive tutorial on using diffusion models for density estimation, focusing on both toy examples and practical applications to EMNIST datasets. It includes various techniques for improving the stability and efficiency of the diffusion process. Key features: - Toy examples demonstrating critical diffusion model concepts - Implementation of ground truth data estimation, error estimation, and score estimation - Techniques for improving stability and efficiency in the diffusion process - Generation of EMNIST data using U-nets and JAX - Colab notebooks for interactive experimentation Tech stack: Jupyter Notebook Why it matters: This tutorial offers a practical guide for developers to understand and implement diffusion models, making it essential for anyone looking to enhance their knowledge in this field. GitHub stars: 67 Repository: varun-ml/diffusion-models-tutorial

Overview: Spin is an official PyTorch implementation of the Speaker-invariant Clustering (Spin) method presented in the Interspeech 2023 paper. It improves pre-trained networks by disentangling speaker information and preserving content representations through speaker-invariant clustering. Key features: - Speaker-invariant Clustering - Fine-tuning on a single GPU for 45 minutes - Improves pre-trained networks and outperforms prior methods in speech recognition and acoustic unit discovery Tech stack: Python Why it matters: Spin is a novel self-supervised learning method that enhances pre-trained models by effectively separating speaker information from content representations, leading to improved performance in speech recognition and acoustic unit discovery tasks.

Overview: AutoMem MCP is a powerful tool that integrates with various AI platforms, ensuring that your AI assistant remembers everything and across all sessions. It uses AutoMem's persistent graph-vector memory service to achieve perfect recall and cross-device sync. Key features: - Cross-device sync across all MCP platforms - Works with Claude Desktop, Cursor IDE, Claude Code, OpenAI Codex, OpenClaw, Hermes, Grok Build, Google Antigravity, and any MCP-compatible AI platform - Persistent memory across sessions, ensuring decisions, patterns, and context are remembered forever - Graph-Vector architecture with 11 public authorable relationship types and sub-second retrieval Tech stack: TypeScript Why it matters: AutoMem MCP provides a seamless and durable memory solution for AI assistants, ensuring consistent and accurate recall across different AI tools and devices. Repository: verygoodplugins/mcp-automem Primary language: TypeScript GitHub stars: 66 GitHub description: MCP client for AutoMem — give Claude, Cursor, Codex, and other MCP tools durable graph+vector memory across conversations.

Overview: MagicWorld is an interactive video world model designed for exploring dynamic scenes using continuous keyboard actions. It achieves 18 FPS on the L40S GPU and better results on the RealWM120K-Val dataset. Key features: - Motion Drift Constraint: Ensures temporal coherence in dynamic regions to prevent motion drift. - Long-Horizon Stability: Preserves historical scene states during autoregressive rollout and improves long-horizon stability. - RealWM120K Dataset: Provides diverse citywalk videos and multimodal annotations for real-world video world modeling. Tech stack: Python Why it matters: MagicWorld is a significant advancement in interactive video world exploration, offering improved stability and realism in dynamic scene modeling.

### Overview: The Chat With RTX Python API provides a powerful tool for developers to integrate real-time text-to-speech functionality into their applications using the RTX engine. ### Key features: - Integrate real-time text-to-speech capabilities - Use RTX engine for efficient and high-performance LLMs - Compatible with Python, making it easy to integrate into existing projects ### Tech stack: Python ### Why it matters: This API streamlines the process of adding voice functionality to applications, enhancing user experience by providing a seamless and engaging conversational interface.

Overview: Seedream 5.0 Pro API is a comprehensive image generation service offering 1K and 2K layer pricing, complete with OpenAI-compatible endpoints and real-time billing. It supports multiple resolutions and layers, making it ideal for various creative projects. Key features: - 1K and 2K image generation at $0.0146 per image - 1K and 2K image generation at $0.0293 per image - Layered pricing with 1K and 2K options - OpenAI-compatible endpoint for seamless integration - Real-time billing and refund on failed tasks - Async processing with task tracking Tech stack: Python Why it matters: Seedream 5.0 Pro API is a cost-effective and efficient solution for generating images with precise text rendering and multi-reference editing. Its real-time billing and OpenAI-compatible endpoint make it ideal for developers and creatives looking to streamline their workflow and save on costs. Repository: ydfqphoks5738xo/seedream5.0-seedream-5.0-gateway Primary language: Python GitHub stars: 67 GitHub description: Seedream 5.0 Pro API (`seedream-5-0-pro`) Seedream 5.0 Pro API is a comprehensive image generation service offering 1K and 2K layer pricing, complete with OpenAI-compatible endpoints and real-time billing. It supports multiple resolutions and layers, making it ideal for various creative projects. Key features: - 1K

### Overview: This curated list of **95 oral papers** from CVPR 2025 covers a wide range of topics in computer vision, including 3D geometry and reconstruction, rendering and 3D/4D generation, humans and avatars, image and video generation, vision-language and video understanding, embodied AI and autonomous driving, segmentation and detection, computational imaging and restoration, learning, efficiency, and trustworthiness, and medical and scientific vision. ### Key features: - **3D Geometry & Reconstruction**: Papers such as VGGT, CUT3R, MoGe, FoundationStereo, Multi-view Reconstruction, MV-DUSt3R+, and Stereo4D. - **Rendering & 3D/4D Generation**: Projects and codes related to rendering, 3D/4D generation, and continuous 3D perception. - **Humans, Avatars & Motion**: Papers like MoGe, FoundationStereo, and Multi-view Reconstruction. - **Image & Video Generation**: Projects and codes for image and video generation. - **Vision-Language & Video Understanding**: Papers and projects related to vision-language and video understanding. - **Embodied AI & Autonomous Driving**: Papers and projects on embodied AI and autonomous driving. - **Segmentation & Detection**: Papers and projects related to segmentation and detection. - **Computational Imaging & Restoration**: Projects and codes for computational imaging and restoration. - **Learning, Efficiency & Trustworthiness**: Papers and projects on learning, efficiency, and trustworthiness. - **Medical & Scientific Vision**: Papers and projects on medical

Overview: PatentRadar is a patent infringement analysis system that provides a quick and efficient way to generate a comprehensive report on potential infringing competitors. By integrating multiple search tools, leveraging advanced machine learning models, and incorporating robust evidence gathering and verification processes, PatentRadar significantly reduces the time and effort required for legal analysis. Key features: - **Input Patent Number**: Accepts a patent public number as input. - **Competitor Analysis**: Generates a detailed report comparing the patent's features with those of potential infringers. - **Evidence Collection**: Extracts relevant public evidence URLs and provides actionable suggestions for further research. Tech stack: Python Why it matters: PatentRadar streamlines the traditional manual process of patent infringement analysis, enabling lawyers to quickly identify potential infringers and gather evidence within a single hour. This tool is particularly valuable for reducing the time and cost associated with complex legal proceedings, making it an essential resource for patent holders and legal professionals alike.