Projects

Overview: PassLLM is a state-of-the-art password guessing AI tool that significantly outperforms other models, achieving up to 45% higher success rates in most scenarios. Leveraging Personally Identifiable Information (PII) and LoRA fine-tuning, it efficiently guesses passwords within 100 guesses, making it a powerful tool for ethical hacking and security assessments. Key features: - State-of-the-Art Accuracy: Achieves +45% higher success rates than leading benchmarks. - PII Inference: Successfully guesses 12.5% - 31.6% of typical users within 100 guesses. - Efficient Fine-Tuning: Utilizes LoRA to lower VRAM usage without sacrificing model reasoning capabilities. - Advanced Inference: Implements the paper's algorithm to prioritize the most likely candidates. - Data-Driven: Trains on millions of real-world credentials to learn deep statistical patterns. - Pre-trained Weights: Includes robust models pre-trained on major breaches and the COMB dataset. Tech stack: Python, PyTorch, CUDA, LoRA, and transformers. Why it matters: PassLLM is a groundbreaking tool that enhances the accuracy and efficiency of password guessing, offering a significant advantage in ethical hacking and security assessments. Its ability to outperform existing models by over 45% and its use of PII for targeted guessing makes it a valuable asset in the field of cybersecurity. Repository: Tzohar/PassLLM GitHub stars: 129 GitHub description: The world's most accurate password guessing

### Overview GenZ is designed to simplify the relationship between the hardware platform used for serving Large Language Models (LLMs) and inference serving metrics like latency and memory. Running an LLM on hardware has three key components: the model, the use case, and the optimization strategies. GenZ can generate latency and memory usage estimations based on the specified LLM, hardware platform, and parallelism configurations. ### Key Features - **Model**: The LLM model architecture and parameters. - **Usecase**: Size of input queries, expected output size, and number of parallel beams. - **Optimization**: Quantization, batching, parallelization, and operator fusion. ### Tech Stack - Jupyter Notebook ### Why It Matters GenZ can help computer architects understand trends in hardware and LLM performance, enabling better design choices for future AI platforms. It can answer system-level questions about deployment platforms, performance optimization, and resource allocation, guiding architects in designing more efficient AI infrastructure. ### GitHub Stars 128 ### Primary Language Jupyter Notebook ### Useful Links - [GenZ-LLM-Analyzer](https://genz-llm-analyzer.streamlit.app/) - [Installation](#installation) - [Examples](#examples) - [Parallelism Scheme](#parallelism-scheme) - [Communication](#communication) - [Data Types](#data-types) - [TODOs](#todos) - [Citation](#citation)

### Overview: Bokeh Diffusion is an innovative project that introduces a method for precise, scene-consistent defocus blur control in text-to-image diffusion models. This technique enables the generation of high-quality images with realistic defocus effects, enhancing the realism and visual appeal of generated images. ### Key Features: - **Scene Consistency**: Bokeh Diffusion ensures that the defocus blur transitions are consistent with the overall scene, providing a more natural and immersive experience. - **Text-to-Image Generation**: The project leverages text-based prompts to generate images, making it versatile for various creative applications. - **Defocus Blur Control**: It allows for precise control over the defocus blur effect, enabling users to achieve desired visual effects. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. - **torch**: Used for deep learning models, particularly in the context of diffusion models. - **flash-attn**: A library for efficient attention mechanisms in neural networks. ### Why It Matters: Bokeh Diffusion offers a significant improvement in the quality and realism of generated images, making it a valuable tool for applications such as image generation, art, and creative design. The ability to control defocus blur effectively enhances the overall visual appeal and realism of the generated images, making it a standout feature in the field of text-to-image diffusion models. ### Repository: The project is hosted on GitHub at <https://github.com/atfortes/BokehDiffusion>. It has been starred by 129 developers, indicating its popularity and utility in the community.

### Overview: LLM Wiki Memory is a local, git-versioned memory system designed for AI coding agents. It captures, compiles, and recalls conversations, ensuring your agent learns from its mistakes and improves over time. Unlike traditional methods, it eliminates the need for external services or RAG stacks, making it a more efficient and reliable solution. ### Key Features: - **Automatic Memory Capture**: Your agent remembers every session, learning from its mistakes. - **Durable Project Knowledge**: Conversations are compiled into durable project knowledge, ensuring your agent applies the right context next time. - **Local MCP Server**: Recalls the right context through a local MCP server, providing offline consolidation while you sleep. ### Tech Stack: - JavaScript ### Why It Matters: LLM Wiki Memory fixes the limitations of traditional methods, ensuring your AI agent never starts from zero again. It captures, compiles, and recalls conversations, providing a more efficient and reliable memory system for coding agents.

Overview: PERSIST is a world model that generates coherent rollouts over thousands of steps by modeling a dynamic 3D state of the world. It uses a voxel-based latent diffusion approach to predict future voxel world states, pixel observations, and camera states based on player actions. Key features: - **Voxel-based Latent-Diffusion World Model**: Generates a dynamic and interactive voxel world. - **Autoregressive Predictions**: Predicts future states based on input player actions. - **Multi-Stage Pipeline**: Uses a denoiser to predict voxel, camera, and pixel states. - **Two VAEs**: Maps voxels/pixels to and from latent space. Tech stack: Python Why it matters: PERSIST is significant for its ability to generate coherent and dynamic world models, making it useful for applications requiring interactive and explorable environments.

### Overview: Welcome to the **Language AI Engineering Lab** — a comprehensive, structured repository designed to guide you from **human language fundamentals and NLP** through **Transformers, Large Language Models**, and into **production-ready Language AI systems**. ### Key Features: - **Human Language Fundamentals and NLP**: Tokenization, embeddings, intent extraction, entity recognition, and text generation. - **Transformer Architecture**: Deep dive into transformers, including attention, embeddings, positional encoding, and why transformers work. - **LLM Fundamentals**: Core LLM concepts, architectures, training strategies (fine-tuning, RLHF), and evaluation foundations. - **Prompt Engineering**: Zero-shot, one-shot, few-shot prompting, reasoning patterns, prompt templates, and optimization techniques. - **Context Management**: Managing LLM context windows, conversation state, memory, truncation strategies, and structured outputs. - **Retrieval-Augmented Generation (RAG) Pipeline**: End-to-end Retrieval-Augmented Generation pipelines: indexing, chunking, embedding, retrieval, reranking, grounding, and response synthesis. - **Context Engineering**: Designing context as a system: instruction hierarchies, memory fusion, grounding strategies, safety constraints, and cost-aware assembly. - **Evaluation and Benchmarks**: Metrics, prompt testing, regression testing, hallucination measurement, latency, cost tracking, and tracing. - **Harness Engineering**: LLM test harness design: scaffolding, input/output capture, retry logic, mocking, CI/CD integration, and automated evaluation pipelines. - **Hallucinations and Fact

### Overview: Bookmarks AI is a tool designed to organize and record your browser bookmarks using AI. It automates the process of categorizing and structuring your bookmarks, making them easily accessible and searchable. ### Key Features: 1. **AI-Driven Bookmark Organization**: Uses GPT-4 to analyze and categorize your bookmarks, ensuring they are organized and easy to find. 2. **Browser Integration**: Syncs your bookmarks with your browser, ensuring your data is secure and accessible. 3. **Personalized AI**: Creates a personalized AI assistant tailored to your browsing habits, enhancing your experience. ### Tech Stack: TypeScript ### Why It Matters: Bookmarks AI simplifies the process of managing and organizing your browser bookmarks, making it easier to find the information you need quickly and efficiently. By automating the categorization and structuring of your bookmarks, it saves you time and effort, allowing you to focus on more important tasks.

Overview: This project presents a novel method for fine-grained control of motion generation through hierarchical semantic graphs, leveraging motion descriptions to achieve comprehensive understanding and precise motion control. Key features: - Hierarchical semantic graphs for motion description disentanglement - Text-to-motion diffusion process decomposition into three semantic levels - Continuous refinement of motion results through edge weights and node modifications Tech stack: Python Why it matters: This approach offers a more nuanced and controllable way of generating motion, making it particularly useful in applications requiring precise control over motion details, such as in animation or simulation tasks. Repository: jpthu17/GraphMotion Primary language: Python GitHub stars: 129 GitHub description: [NeurIPS 2023] Act As You Wish: Fine-grained Control of Motion Diffusion Model with Hierarchical Semantic Graphs

### Overview: Oído is a speech recognition system designed for ESP32-S3, achieving 3.7% LibriSpeech WER with no cloud or neural accelerator. Built by Lokutor, it is a compact solution for real-time speech recognition. ### Key Features: - **Compact ESP32-S3 Chip**: Runs on a $5 ESP32-S3-WROOM-1-N16R8 board. - **No Cloud, No NPU**: Achieves 3.7% LibriSpeech WER with no cloud or neural accelerator. - **Real-Time Speech-to-Text**: Runs on a single core at 240 MHz, with English and Spanish models available in streaming and utterance modes. - **Transcripts**: Transcripts are identical to the instruction-exact emulator on all 27 clips, with accuracy numbers from the host build of the same C code. ### Tech Stack: - **Lokutor**: Built by Lokutor. - **Models**: Models on Hugging Face: [int8](https://huggingface.co/lokutor-ai/oido-ctc-small-int8) · [int4](https://huggingface.co/lokutor-ai/oido-ctc-small-int4). ### Why It Matters: - **Compact and Affordable**: Ideal for IoT and small embedded systems. - **Real-Time Performance**: Achieves real-time accuracy with minimal latency, making it perfect for voice agents and command-line interfaces. - **Versatile Languages**: Supports both English and Spanish, with streaming and

Overview: This repository provides an easy-to-use workspace for fine-tuning the Qwen3-TTS model, streamlining the process from raw audio ingestion to high-quality multilingual speech synthesis. Key features: - Integrated Pipeline: Automated audio splitting, ASR transcription, dataset cleaning, and tokenization. - Modern WebUI: Premium Gradio interface for data preparation, training monitoring, and interactive inference. - Robust CLI: Unified command-line interface for professional automation and remote server management. - Optimized Presets: Hardcoded, expert-tuned training parameters for different model variants (0.6B / 1.7B). - Docker Ready: Out-of-the-box environment support via pre-configured Docker images. Tech stack: Python, PyTorch. Why it matters: This repository offers a streamlined workflow for fine-tuning the Qwen3-TTS model, ensuring timbre stability, expressive control, and cross-lingual synthesis without the need for zero-shot fine-tuning. Repository: mozi1924/Qwen3-TTS-EasyFinetuning GitHub stars: 129 GitHub description: Easy fine-tuning for Qwen3-TTS: Fast voice cloning and high-quality multilingual speech synthesis.

Overview: OpenLegion is a secure, self-hosted AI agent runtime designed for teams that can't afford a security incident. It runs fleets of autonomous agents in isolated Docker containers, with per-agent budgets and a central credential vault for API calls. Key features: - Secure, self-hosted AI agent runtime - Runs fleets of autonomous agents in isolated Docker containers - Per-agent budgets and per-agent spend limits - Central credential vault for API calls - Trusted mesh host for coordination - Source-available under PolyForm Perimeter License 1.0.1 Tech stack: Python Why it matters: OpenLegion provides a secure, self-hosted alternative to OpenClaw, offering source-code auditability and per-agent budgeting, making it ideal for teams that prioritize security and autonomy. Repository: openlegion-ai/openlegion GitHub stars: 127

Overview: The project aims to develop a bilingual Urdu-English retrieval-augmented generation system for government and legal documents in Pakistan, using OCR, hybrid retrieval, citation grounding, and evaluation. Key features: - Retrieval of relevant passages - Generation of answers only from retrieved passages - Citation of the source page - Evaluation of answers with a focus on failures Tech stack: Python Why it matters: The project is a research study on the application of retrieval-augmented generation in low-resource languages, specifically focusing on the Urdu-English hybrid retrieval system for the Constitution of Pakistan. Repository: salikhussain71-code/PAKGOV-RAG-project Primary language: Python GitHub stars: 129 GitHub description: Research project on bilingual Urdu-English retrieval-augmented generation for Pakistani government and legal documents, with OCR, hybrid retrieval, citation grounding, and evaluation.

Overview: pyannote-rs is a Rust implementation of Pyannote, a powerful audio diarization tool, designed for efficient and accurate speaker identification in audio recordings. Key features: - Fast performance with DirectML on Windows and CoreML on macOS. - Accurate timestamps with Pyannote segmentation. - Identifies speakers using Wespeaker embeddings. Tech stack: Rust Why it matters: pyannote-rs offers a fast and accurate solution for speaker diarization, making it ideal for real-time applications and high-performance audio processing. Repository: thewh1teagle/pyannote-rs Primary language: Rust GitHub stars: 129

Overview: Timescale Vector Cookbook is a collection of recipes and sample application starter kits designed to help developers build applications with Large Language Models (LLMs) using PostgreSQL and Timescale Vector. Key features: - Timescale Vector Recipes - PostgreSQL and Timescale Vector Integration - Timescale Vector Python Client Library Tech stack: Jupyter Notebook Why it matters: This repository provides practical solutions and examples for developers looking to integrate LLMs with PostgreSQL and Timescale Vector, enhancing their applications with vector, relational, and time-series data capabilities. Repository: timescale/vector-cookbook GitHub stars: 129

### Overview: Scale-wise Distillation (SwD) accelerates diffusion models by turning them into progressive few-step models, achieving up to 2x speedup compared to full-resolution few-step alternatives while maintaining or improving image quality. ### Key features: - Scale-wise Distillation (SwD) framework for accelerating diffusion models. - Four versions of SwD available: `SDXL-SwD (2.6B)`, `SD3.5-Medium-SwD (2.2B)`, `SD3.5-Large-SwD (8B)`, and `FLUX-SwD (12B)`. - Requires two key arguments: `scales` and `sigmas`. - `scales` defines a sequence of spatial latent resolutions. - `sigmas` corresponds to the few-step timestep schedule. ### Tech stack: Python ### Why it matters: SwD significantly speeds up the inference process of diffusion models, making them more efficient and practical for real-time applications and high-resolution image generation.

### Overview: ReDi is a groundbreaking project that introduces a novel approach to enhancing generative image modeling. By leveraging joint image-feature synthesis, ReDi significantly boosts the quality and diversity of generated images, making it a standout contribution to the NeurIPS 2025 Spotlight. ### Key Features: - **Joint Image-Feature Synthesis**: ReDi combines image synthesis and feature learning to create more realistic and detailed images. - **Boosting Generative Models**: It improves the performance of generative models, particularly in terms of image quality and diversity. - **NeurIPS 2025 Spotlight**: ReDi was accepted as a Spotlight paper, highlighting its significance in the field. ### Tech Stack: The project is implemented using Python, leveraging the PyTorch framework for deep learning and the Hugging Face library for model training and evaluation. ### Why It Matters: ReDi represents a significant advancement in generative image modeling, offering improved results over existing techniques. Its ability to enhance the quality and diversity of generated images makes it a valuable tool for researchers and practitioners in the field. ### Repository: The repository is hosted on GitHub at [zelaki/ReDi](https://github.com/zelaki/ReDi). It has received 129 GitHub stars, indicating its popularity and impact in the community. ### Setup: To get started with ReDi, clone the repository and set up the necessary environment: ```bash git clone https://github.com/zelaki/ReDi.git cd ReDi conda env create -f environment.yml conda activate ReDi

### Overview: MioVRC Translator 是一款专为VRChat和VRC用户设计的实时语音翻译工具，旨在提供即时的语音翻译体验。它支持多种语言的文本翻译，并具备反向翻译功能，帮助用户在VR环境中更好地理解和交流。 ### Key features: - **本地语音识别与实时翻译**：通过麦克风捕捉语音，识别文本并实时翻译，随后发送到VRChat的Chatbox。 - **反向翻译功能**：监听VRChat中的音频并进行反向翻译，方便用户确认和理解对话内容。 - **同声传译功能**：将翻译文字通过AI语音朗读，并通过MixLine虚拟麦克风让VRChat中的其他玩家听到。 - **多种输出格式支持**：支持多种聊天框输出格式，方便用户根据需求选择。 - **多语言界面**：提供多种语言界面，适应不同用户需求。 - **自定义选项**：支持多种自定义选项，如麦克风设置、输出格式选择等。 - **模型缓存**：支持自定义模型缓存目录，方便用户根据需求选择。 ### Tech stack: Python ### Why it matters: MioVRC Translator为VRChat和VRC用户提供了高效、便捷的语音翻译工具，极大地提升了他们的沟通体验。通过实时翻译和反向翻译功能，用户可以轻松地在VR环境中进行交流，无需担心语言障碍。此外，同声传译功能使得多人对话更加流畅，增强了团队协作的效率。

### Overview: A deep-thinking RAG pipeline is designed to handle complex queries by integrating reasoning, reflection, and smart decision-making, unlike a simple linear approach. The pipeline decomposes queries into structured research plans, uses adaptive retrieval strategies, and synthesizes evidence into a comprehensive answer. ### Key features: - **Decomposition of Queries**: The pipeline breaks down complex queries into structured research plans. - **Adaptive Retrieval**: It uses a supervisor to dynamically choose the best search strategy (vector, keyword, or hybrid) for each step. - **Reranking and Compression**: A high-precision cross-encoder reranks initial results and a distiller agent compresses the best evidence. - **Cumulative Understanding**: After each step, the agent summarizes findings and updates its research history. - **Strategic Decision Making**: A policy agent inspects the research history and makes decisions to continue, revise, or finish the research. ### Tech stack: Jupyter Notebook ### Why it matters: The deep-thinking RAG pipeline demonstrates how to handle complex, multi-hop queries by integrating reasoning and reflection, making it more effective than a simple linear approach.

### Overview: This project implements the ICML 2023 paper "Specializing Smaller Language Models towards Multi-Step Reasoning" by Yao Fu, Hao Peng, Litu Ou, Ashish Sabharwal, and Tushar Khot. The goal is to specialize smaller language models (FlanT5) towards multi-step reasoning tasks, focusing on in-context and zero-shot capabilities. ### Key Features: - **Data Engineering**: The project includes detailed data engineering efforts, processing data into four formats: in-context answer-only, in-context chain-of-thought, zero-shot answer-only, and zero-shot chain-of-thought. - **Training Script**: The primary training script is `train_distill_simple.py`, which is straightforward and focuses on data engineering and hyperparameter tuning. - **Model Usage**: The project uses FlanT5 as the base model, and quickstart scripts are provided for running the model with specific parameters. ### Tech Stack: - **Language**: Jupyter Notebook ### Why It Matters: This project is significant for researchers and developers interested in improving the performance of smaller language models (LLMs) in multi-step reasoning tasks. By specializing these models towards specific reasoning capabilities, it can enhance their utility in various applications, such as code completion, question answering, and more. ### Repository: - **Repository**: FranxYao/FlanT5-CoT-Specialization - **GitHub Stars**: 128 - **GitHub Description**: Implementation of ICML 2023 Paper: Specializing Smaller Language Models towards Multi

### Overview: This project is an official implementation of "Using Diffusion Priors for Video Amodal Segmentation" at CVPR 2025. The code is designed to provide checkpoints and inference capabilities for various modalities, including depth, and fine-tuning for specific datasets like SAIL-VOS and TAO-Amodal. ### Key features: - **Model Checkpoints**: Available for SAIL-VOS and TAO-Amodal. - **Fine-tuning**: Support for fine-tuning Diffusion-VAS on different datasets. - **Inference**: Runs inference on birdcage examples and can be customized with different sequences. - **Custom Data**: Supports downloading and using custom video data for inference. ### Tech stack: Python ### Why it matters: This project provides a comprehensive solution for video amodal segmentation, leveraging diffusion priors to enhance segmentation accuracy. The availability of checkpoints and fine-tuning capabilities makes it a valuable resource for researchers and developers working on similar tasks.