Projects

### Overview: Controllable Video Generation with Diffusion Models is an awesome project that leverages diffusion models to generate videos with precise control over various aspects such as pose, audio, expression, and more. This project is highly valuable for anyone looking to enhance the quality and flexibility of video content. ### Key Features: - **Pose Control**: Enables precise control over human poses using UniAnimate-DiT. - **Audio Control**: Allows for the generation of videos with controlled audio using EchoMimicV2. - **Expression Control**: Offers the ability to modify facial expressions with MikuDance. - **Universal Control**: Provides a flexible framework for generating videos with controlled parameters. - **Camera Control**: Allows for the manipulation of camera angles and perspectives. - **Trajectory Control**: Enables the generation of videos with controlled motion paths. - **Subject Control**: Controls the focus and perspective of the subject in the video. - **Area Control**: Allows for the creation of videos with controlled areas of interest. - **Video Control**: Offers precise control over video content, including color and style. - **Brain Control**: Enables the generation of videos with controlled brain activity. - **ID Control**: Controls the identity of the subject in the video. ### Tech Stack: The project utilizes diffusion models, specifically those from the UniAnimate-DiT, EchoMimicV2, MikuDance, and DreaMoving frameworks. ### Why It Matters: This project is highly valuable for anyone looking to enhance the quality and flexibility of video content. It provides a powerful toolset for generating videos with precise control over

### Overview: This repository focuses on a collection of awesome 3D generative models, bridging the gap between text, images, and 3D data. It includes models for 3D object generation, texture generation, and scene modeling. ### Key Features: - **Modeling - 3D Representation**: Models 3D shapes using techniques like VecSet and TRELLIS. - **Modeling - VAE for Mesh Reconstruction**: Uses Variational Autoencoders to reconstruct 3D meshes. - **Modeling / Texture - Flow-matching DiT for Mesh Generation & Texture Generation**: Applies flow-matching techniques to generate textures and meshes. - **Modeling - Re-topology / Artist Mesh Generation**: Reconstructs 3D meshes from artist-generated meshes. - **Modeling - UV Unwrapping**: Unwraps 3D meshes for better texture mapping. - **Texture Generation**: Generates textures for 3D models. - **Motion - Rigging / Articulation / Dynamics with Physical Property / Animation / Simulation**: Manages 3D motion and dynamics. - **Acceleration For Mesh Generation**: Enhances mesh generation processes for better performance. - **Post-train of 3D base models**: Includes post-training techniques for improved 3D models. - **3D Scene Generation (World Models)**: Covers both open-source and closed-source models for creating 3D scenes. ### Tech Stack: None explicitly mentioned. ### Why It Matters: This repository provides a comprehensive collection of tools and techniques for generating 3D models, textures, and

Overview: Vosk is a cloud-free speech recognition toolkit with high accuracy and low latency, designed for real-time applications. Key features: - Local speech recognition without waiting for the end of a sentence. - Supports 16 languages with pre-trained models. - Easy to use with Rust, providing a simple API for developers. Tech stack: Rust Why it matters: Vosk is highly accurate and efficient, making it suitable for applications requiring real-time speech recognition without the need for cloud services, thus saving on latency and bandwidth costs.

Overview: This repository is a template designed to efficiently fine-tune any HuggingFace Large Language Model using distributed training on multiple GPUs and leveraging Ray AIR for cluster orchestration and DeepSpeed for parameter and optimizer sharding. Key features: - Distributed training on multiple GPUs - Ray AIR for cluster management - DeepSpeed for parameter and optimizer sharding - Fine-tuning of FALCON-7B model Tech stack: Python Why it matters: This setup allows for efficient fine-tuning of large language models by leveraging distributed training and advanced optimization techniques, making it suitable for large-scale and high-performance computing environments.

### Overview: Real Agents is a Unity-based planning framework that combines traditional AI methods with Large Language Models (LLMs) to create generative AI agents. This project aims to address the challenges of implementing AI in games by providing a robust and efficient solution. ### Key Features: - **Goal Oriented Action Planning (GOAP) Integration**: Real Agents seamlessly integrates GOAP to ensure the agent's actions are directly based on predefined behaviors. - **Behavior Execution and Status Monitoring**: The agent's behavior is executed and monitored by the program, while planning is triggered only when the status changes. - **Large Language Model (LLM) Based AI**: The LLM is used to generate plans, making it easier to debug and faster in response times compared to relying solely on LLMs. ### Tech Stack: Unity, C#, GOAP, LLMs, and AI Frameworks. ### Why It Matters: Real Agents provides a practical solution for integrating AI in games, addressing the difficulties of debugging and slow response times commonly encountered when using LLMs. By combining traditional AI methods with LLMs, it offers a more efficient and easier-to-manage approach to creating generative AI agents in games.

Overview: ADvISER is a flexible framework designed to facilitate task-oriented dialog system research and development. Key features: - Supports text-to-text conversation - Enables speech synthesis and speech recognition - Provides a modular architecture for easy integration of different components - Offers a comprehensive documentation and installation guide Tech stack: Python Why it matters: ADvISER simplifies the development process for researchers and developers working on task-oriented dialog systems, making it easier to create and test various conversational models. Repository: DigitalPhonetics/adviser GitHub stars: 60 GitHub description: ADvISER is a flexible framework to encourage task-oriented dialog system research & development.

### Overview: GcrawlAI is a powerful, enterprise-grade web crawler and scraper platform designed to feed retrieval-augmented generation (RAG) pipelines, LLMs, and semantic search indexes. It converts complex, noisy web structures into clean Markdown, structured JSON metadata, and full-page screenshots, making it ideal for AI applications. ### Key Features: - **🥷 Fingerprint Hygiene & Stealth Browsing**: Masks automated runtimes, WebGL signatures, canvas fingerprints, and automation leaks to bypass aggressive anti-bot protections. - **🔀 Stepped Residential Proxy Rotation**: Multi-tier automatic proxy escalation with geographic IP targeting to match the target site's local region. - **✨ Fit-Markdown Extraction**: Converts pages to clean, LLM-ready markdown, pruning HTML boilerplate, menus, footers, and advertisements. - **💾 Offline HTML Bundle**: Downloads full pages along with CSS, images, and other assets, packaged into a single ZIP file for local offline rendering. Available as a REST API endpoint and browser extensions. - **🔐 Extension-based Cookie Synchronization**: Extracts active session cookies and local storage tokens directly from your Chrome/Firefox browser using the official GcrawlAI browser extension. Automatically syncs them with the backend to scrape login-required websites without re-authentication. - **📊 SEO Data Collection**: Automatically extracts metadata, headers, titles, descriptions, open graph tags, and links structure from crawled pages. - **📸 High-Resolution Screenshotting & Document Parsing**: Uses physics-based scrolling to capture lazy-loaded content correctly. - **🗺️ URL Mapping**: `/links` endpoint

Overview: ThinkJEPA is a dual-path framework that combines a vision-language model (VLM) with a JEPA branch to create a high-level reasoning cortex and low-level prediction cerebellum for embodied prediction tasks. Key features: - VLM-thinker branch: Provides high-level reasoning guidance from Qwen3-VL-Thinking features. - Dense JEPA branch: Models video dynamics and supplies low-level embodied features for prediction. - Released training path: Predicts future trajectory outputs from JEPA features conditioned by pyramid guidance from the VLM branch. Tech stack: Python Why it matters: ThinkJEPA offers a novel approach to embodied prediction by leveraging a vision-language model to guide high-level reasoning and a JEPA branch to predict low-level dynamics, making it suitable for applications requiring both semantic understanding and physical consistency. Repository: Hai-chao-Zhang/ThinkJEPA GitHub stars: 60 GitHub description: [NeurIPS 2026] ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model License: Not specified

Overview: Wav2vec 2.0 is a self-supervised learning framework implemented in Pytorch, as described in the paper "wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations". Key features: - Implementation of Wav2vec 2.0, a state-of-the-art model for speech representation learning. - Utilizes self-supervised learning techniques to enhance model performance. - Supports both training and inference on various speech datasets. Tech stack: Pytorch Why it matters: Wav2vec 2.0 is a significant advancement in speech representation learning, offering improved performance and efficiency compared to traditional methods. Its self-supervised learning approach allows for effective model generalization across diverse speech tasks, making it a valuable tool in the field of speech processing and natural language processing.

### Overview: In this project, the authors propose a novel method called "Reliable AD" to generate high-quality advertising images. By integrating human feedback into the generative process, the system ensures that the generated images are reliable and suitable for advertising, thereby reducing the need for extensive manual inspection and improving production efficiency. ### Key Features: - **Reliable Feedback Network (RFNet)**: Automatically evaluates the quality of generated images based on human feedback. - **Recurrent Generation (Recurrent Generation)**: Enhances the number of available advertising images by incorporating feedback from RFNet. - **Consistent Condition Regularization (RFFT)**: Fine-tunes diffusion models using feedback from RFNet to produce highly efficient and visually appealing images. - **RF1M Dataset**: A large-scale dataset of over 1 million annotated images, which is crucial for training RFNet to accurately assess image availability and provide human-like feedback. ### Tech Stack: The project is implemented using Python, with the primary language being Python 3.9 and the use of PyTorch 2.0.1 for training and inference. ### Why It Matters: The approach offers a reliable solution for advertising image generation, significantly reducing the need for extensive manual inspection and improving production efficiency. The inclusion of human feedback in the generation process ensures that the generated images are not only visually appealing but also suitable for advertising purposes, thereby increasing the rate of available generated images.

### Overview: Auto-Rubric as Reward is a project that aims to bridge the gap between implicit preferences and explicit, inspectable rubric criteria for multimodal tasks. It provides a compact implementation of Auto-Rubric as Reward, enabling users to generate, verify, and refine rubrics for a wide range of visual generation and image editing tasks. The project focuses on generating rubrics from a set of labeled visual preference examples, using a frozen VLM judge to produce pairwise rewards for Rapid Prototyping of Objectives (RPO). ### Key Features: - **Explicit Reward Criteria**: The "reward model" is readable rubric text rather than a hidden scalar model. - **Verifiable Generation Loop**: Candidate rubrics are checked against labeled examples and revised when they fail. - **Pairwise Visual Rewards**: Rank 1 receives `1.0`; rank 2 receives `-0.1` for RPO. - **T2I and Edit Support**: Prompt-only FLUX and source-image-aware Qwen-Image-Edit paths are both wired. - **Reusable Rubric Files**: Generate rubrics once, inspect them, and reuse the same file for deterministic training launches. - **OpenAI-compatible VLMs**: Use local Qwen3-VL through vLLM or host. ### Tech Stack: - Python ### Why It Matters: Auto-Rubric as Reward provides a practical solution for generating and evaluating rubrics in multimodal learning tasks, making it easier to align implicit preferences with explicit criteria. This project contributes to the growing interest in rubric-based reward learning and opens

Overview: Percolate is a Postgres extension designed for AI workflows and agent memory, enabling SQL-driven workflows with agent memory built from conversations and documents. Key features: - AI-driven incident desk workflow - Durable workflows and agents in SQL - Agent memory built from conversations and documents - Dated memory for querying as of any day - Advanced memory features for answering questions based on remembered facts - Workflow steps and query modes in one place Tech stack: Percolate is built on PostgreSQL and Rust, utilizing SQL for defining, starting, approving, querying, and inspecting workflows. Why it matters: Percolate simplifies the integration of AI workflows and agent memory into Postgres, providing a powerful toolset for developers and data engineers to manage and analyze complex AI-driven processes.

### Overview: CTC-OptimizedLoss is a Python library that optimizes the Minimum Word Error Rate (WER) Loss for Connectionist Temporal Classification (CTC) with beam search. It includes implementations for TensorFlow, PyTorch, and is particularly useful for knowledge distillation and label smoothing in CTC loss. ### Key features: - **TensorFlow Implementation**: Provides MWER (minimum WER) Loss with CTC beam search. - **PyTorch Implementation**: Implements Delay-penalized CTC based on Finite State Transducer. - **Knowledge Distillation**: Supports knowledge distillation for CTC loss. - **Label Smoothing**: Offers KL divergence loss for label smoothing. ### Tech stack: Python ### Why it matters: CTC-OptimizedLoss is essential for researchers and developers working on speech recognition and sequence-to-sequence models, offering optimized solutions for CTC loss and knowledge distillation.

### Overview LittleAIBox is a privacy-focused AI conversation platform designed for chat, coding, research, and document analysis. It supports Gemini-compatible conversational AI, client-side parsing for various file types, and dual API-key management for secure access. ### Key features - **Privacy-Preserving**: User-provided keys and local file handling for secure data management. - **Conversational AI**: Supports general, coding, and research tasks. - **Client-Side Parsing**: Parses PDF, DOCX, PPTX, and XLSX files. - **Mermaid, LaTeX, and Vega-Lite Rendering**: Enables rich visual content rendering. - **Conversation Timeline and Presets**: Tracks conversations and task-specific settings. - **Dual API-Key Management**: Health checks and failover for secure access. - **Voice Input**: Allows for customizable model parameters. - **PWA Offline Support and Android Packaging**: Provides offline capabilities and native Android support. - **Optional Web Search and Semantic Scholar Research Mode**: Enhances research capabilities. ### Tech stack - **Frontend**: Vite - **Native Android**: Capacitor - **Relay and Availability**: Cloudflare edge workers - **IndexedDB**: Local conversation state storage - **Model APIs**: Gemini-compatible ### Why it matters LittleAIBox is designed for privacy-conscious teams, developers, researchers, and operators in restricted networks, offering a secure and efficient way to access AI-powered tools without compromising user privacy. ### Repository Repository: [antonkarasbiz/littleaibox](https://github.com/anton

Overview: This project demonstrates how to serve an OCR model using BentoML, a framework for deploying machine learning models as microservices. It accepts images as input and returns the text contained within. Key features: - Serving an OCR model - Accepting images as input - Returning text and bounding box coordinates Tech stack: Python Why it matters: This project serves as an example for developers looking to deploy and manage their machine learning models using BentoML, making it easier to create and scale microservices for real-time applications.

Overview: OverGraph is an absurdly fast embedded graph database with built-in vector search, offering sub-microsecond reads and native connectors for Node.js and Python. It's pure Rust, running on the same engine as Node.js and Python, making it genuinely fast and efficient. Key features: - Sub-microsecond reads - Native connectors for Node.js and Python - Pure Rust core - Built-in vector search - Graph structure and vector similarity in the same engine - AI agent memory, knowledge graphs, RAG pipelines, and semantic search - Built for AI agent memory, knowledge graphs, RAG pipelines, and semantic search - Built for AI agent memory, knowledge graphs, RAG pipelines, and semantic search - Built for AI agent memory, knowledge graphs, RAG pipelines, and semantic search - Built for AI agent memory, knowledge graphs, RAG pipelines, and semantic search - Built for AI agent memory, knowledge graphs, RAG pipelines, and semantic search Tech stack: Pure Rust Why it matters: OverGraph is a graph database that runs inside your process, no server, no network calls, no Docker containers. It's built to feel like a library, not a service you have to operate. It's genuinely fast, not just "fast for a database," but fast enough that you forget it's there.

Overview: Doppelganger is a GitHub App designed to automatically detect and flag duplicate or related issues and pull requests, reducing open-source maintenance time and improving code review efficiency. Key features: - Automatically compares new issues and pull requests against existing ones. - Identifies and comments on highly similar issues to close duplicates. - Provides feedback on PRs based on title and description. - Uses MiniLM-L6-v2 for vector representation and ChromaDB for similarity search. Tech stack: Python, MiniLM-L6-v2, ChromaDB, Ollama. Why it matters: By automating the detection and management of duplicate issues and pull requests, Doppelganger helps open-source maintainers save time and reduce errors, ultimately improving the quality and efficiency of code reviews and project management. Repository: dannyl1u/doppelganger Primary language: Python GitHub stars: 60 GitHub description: Similarity detection for Github Issues + AI code reviews README excerpt: # doppelgänger <ins>Problem:</ins> open-source maintainers spend a lot of time managing duplicate/related (doppelgänger) issues & pull requests <ins>Solution:</ins> doppelgänger compares newly submitted issues/PRs against existing ones, closing and commenting on highly similar issues/PRs **Topics: vector db, github, open-source, embedding search, rag, similarity scores** https://github.com/dannyl1u/doppelganger/assets/45186464/cdc1c68b-

### Overview: Voice ChatBot is a project that combines a voice bot and a RESTful server for interacting with it. It uses PocketSphinx for speech recognition and RHVoice for speech synthesis. The AttentionSeq2Seq model is implemented using Python3+TensorFlow+Keras. ### Key Features: - **Speech Recognition**: Uses PocketSphinx for recognizing speech. - **Speech Synthesis**: Uses RHVoice for generating speech. - **Attention Seq2Seq Model**: Implements an AttentionSeq2Seq model for dialogue interaction. ### Tech Stack: Python3, TensorFlow, Keras ### Why it Matters: This project showcases the integration of speech recognition and synthesis with a neural network model, demonstrating its application in a conversational bot.

Overview: Discus is a Python library designed to generate high-quality data for machine learning and artificial intelligence tasks, leveraging large language models to ensure optimal results. Key features: - Automatically generates user-guided data for ML/AI tasks - Supports LLM-Instances and LLM-Instructions - Provides a seamless integration with popular LLM providers like OpenAI Tech stack: Python Why it matters: Discus simplifies the process of obtaining high-quality datasets, crucial for effective machine learning testing and evaluation, by leveraging the capabilities of large language models to create tailored data sets.

Overview: EuLLM is an open-source platform designed to create, distribute, and run sovereign EU-compliant large language models. It allows verticalizing any model for specific domains, languages, and brands, ensuring compliance with the EU AI Act. Key features: - Decisions in milliseconds, up to 64 questions per page in 0.66 seconds. - A RAG gate that stops when it knows it can't answer. - The same question, the same answer, ensuring consistency. - Decisions that can be audited, corrected, and learned from. - A 125B model on a 16 GB GPU, writing 69 tokens per second. - Up to 62% faster answers with `--mtp`. - A long prompt never freezes the others, with 1.5 seconds for an 8,000-token prompt. - Sixteen users, one GPU, and continuous batching. - Chat models, embedder, and decision model side by side. - Built-in chat, live monitor, and Model hub. Tech stack: Python. Why it matters: EuLLM is designed to be local-first, sovereign, and compliant with the EU AI Act, making it ideal for applications where data sovereignty and privacy are paramount. Repository: eullm/eullm GitHub stars: 60 GitHub description: Open-source platform for creating, distributing and running sovereign EU-compliant LLMs.