Projects

### Overview: ProCyon is an open-source model designed to predict protein phenotypes across various scales. It is built on a multimodal foundation and is available in both a full and a benchmarking version. ### Key Features: - **Full Model**: Provides comprehensive predictions across multiple scales. - **Benchmarking Model**: Optimized for performance and accuracy. - **Dataset Access**: Includes datasets for both training and inference. - **Model Weights**: Available for download from HuggingFace. ### Tech Stack: Python ### Why it Matters: ProCyon offers a powerful tool for researchers and developers to predict protein phenotypes, enhancing the understanding and analysis of protein functions and interactions.

Overview: SmolSQLAgents is a suite of AI agents designed to interact with databases autonomously, providing insights and validated responses. Key features: - Auto Discover Schema - Auto Discover Relationships - Auto Generate Documentation - Natural Language 2 SQL - RAG Search Tech stack: Python Why it matters: SmolSQLAgents simplifies database interactions by offering intelligent, validated responses and seamless integration with enterprise-level databases. Repository: montraydavis/SmolSQLAgents Primary language: Python GitHub stars: 61 GitHub description: The Most Powerful Open-Source Autonomous SQL Agents - RAG, NL2SQL

### Overview: RCDMs, or Rich-Contextual Conditional Diffusion Models, is a novel approach designed to enhance the consistency and coherence of story visualization. By integrating rich contextual conditions and one-pass inference, RCDMs generate stories that are both semantically and temporally consistent, making them a significant advancement in the field of story generation. ### Key Features: - **Rich-Contextual Conditional Diffusion Models (RCDMs)**: A two-stage framework that incorporates rich contextual information at both the image and feature levels to ensure semantic and temporal consistency. - **One-Pass Inference**: RCDMs can generate consistent stories with a single forward inference, improving efficiency and reducing computational complexity. - **Semantic and Temporal Consistency**: RCDMs ensure that stories generated are both semantically and temporally consistent, enhancing the coherence and quality of the visual narratives. ### Tech Stack: - **Python**: The primary programming language used for implementing RCDMs. - **PyTorch**: Used for the implementation of the model architecture and training process. ### Why It Matters: RCDMs significantly improve the consistency and coherence of story visualizations, making them a valuable tool for applications in game development and comic drawing. By addressing the challenges of generating consistent and coherent stories, RCDMs offer a more effective solution compared to traditional methods, particularly in scenarios where temporal and semantic consistency are crucial. ### Repository: muzishen/RCDMs - **Primary Language**: Python - **GitHub Stars**: 62 - **GitHub Description**: RCDMs is a two-stage approach

Overview: pyseekdb is a Python SDK designed for efficient and user-friendly interaction with OceanBase and OceanBase seekdb, providing unified vector, full-text, and hybrid retrieval capabilities. Key features: - Unified API for embedded and remote server modes - Efficient vector similarity search - Hybrid search combining vector and full-text retrieval - Built-in support for various embedding models - Easy collection management and database operations Tech stack: Python Why it matters: pyseekdb simplifies database operations, making it easier and more efficient to manage and query data in OceanBase and OceanBase seekdb, enhancing application workflows. Repository: oceanbase/pyseekdb GitHub stars: 62 GitHub description: A unified python SDK supports OceanBase or OceanBase seekdb, more efficient and easy-to-use.

Overview: DataClaw is a cutting-edge system designed to refine and structure raw multimodal data from unstructured streams, enabling more efficient and effective data acquisition and AI post-training. This project aims to address the limitations of existing passive annotation methods by introducing a novel paradigm called Agentic Data Tailoring. Key features: - **DataClaw-O** (Omni Model): A unified model that can be deployed across multiple domains. - **DataClaw-E** (Experts): A panel of domain-specific models. - **DataClaw-val Benchmark**: The first benchmark dedicated to data refinement, scoring outputs by JSON validity and schema-aware Field / Semantic / Sequence metrics. Tech stack: The project leverages a combination of Hugging Face, Supervised Fine-Tuning, and Group Relative Policy Optimization. Why it matters: DataClaw addresses the challenges of high data entropy and the limitations of existing annotation methods, aiming to unlock the procedural logic embedded in raw data and improve the quality of AI models through refined data processing.

Overview: Vroom Live is a browser-based interface that allows direct communication with your Ableton Live session, enabling you to view and manipulate your tracks, instruments, and devices. Key features: - Real-time chat with Ableton Live - View and modify tracks, instruments, and devices - Modulate parameters using GPT3.5 and OSC tools Tech stack: JavaScript, GPT3.5, OSC, WebSocket Why it matters: Vroom Live simplifies the interaction between web users and their digital audio workstation, enhancing productivity and collaboration in live performance and recording environments. Repository: vroomai/live Primary language: JavaScript GitHub stars: 62 GitHub description: 🎧 Chat with Ableton Live in your browser.

### Overview: ORBIA is a benchmark designed to evaluate the 3D consistency of interactive video world models. It assesses three key dimensions: input preservation, generated-content persistence, and 3D self-consistency. The repository includes evaluation code, model adapters, and data construction pipelines for real videos and Unreal Engine environments, with diverse, curated datasets and six exploration-and-revisit trajectories. ### Key Features: - **Input Preservation**: Ensures the scene and landmarks in the input image are preserved. - **Generated-Content Persistence**: Newly generated content remains consistent when the camera returns to the original view. - **3D Self-Consistency**: Generated views agree in geometry and support a shared 3D reconstruction. - **Exploration-and-Revisit Trajectories**: Six trajectories (forward-backward, lateral, yaw, arc, orbit, peek) move the camera away from the input view, revisit regions outside it, and return to the start, each with a 22-second and a minute-long version with three revisits. ### Tech Stack: - Python ### Why It Matters: ORBIA provides a comprehensive evaluation framework for 3D consistency in interactive video world models, ensuring that generated videos maintain the integrity of the input scene and accurately represent the content across different camera positions. This benchmark is crucial for advancing the field of 3D world modeling and interactive video technology.

Overview: DDSM, a novel framework for optimizing diffusion models, achieves up to 76% reduction in computational costs without compromising on quality. Key features: - Achieves up to 76% reduction in computational costs for diffusion models. - Compatible with latent diffusion. - Compatible with sampling schedulers like DDIM and EDM. Tech stack: Python. Why it matters: DDSM is a significant advancement in the field of diffusion models, offering substantial computational efficiency improvements without sacrificing quality.

Overview: OMatG is a generative framework designed for the prediction and de novo generation of inorganic crystals. It includes crystal structure prediction and *de novo* generation modes, leveraging PyTorch Lightning for ease of use. The framework is accompanied by papers on ICML 2025, NeurIPS 2025, and ICML 2026, providing comprehensive metrics and datasets for evaluation. Key features: - Crystal structure prediction (CSP) - *De novo* generation (DNG) - PyTorch Lightning for command-line interface Tech stack: Python Why it matters: OMatG offers a powerful tool for crystal structure prediction and de novo generation, making it valuable for researchers and practitioners in materials science and related fields. GitHub stars: 61

### Overview: The AI Lead Generation Agent automates the process of discovering and qualifying potential leads from Quora, using Firecrawl for intelligent web scraping, Phidata for agent orchestration, and Composio for Google Sheets integration. This system continuously generates and organizes qualified leads with minimal human intervention, making it ideal for sales and marketing teams to focus on building relationships and closing deals. ### Key features: - **🎯 Targeted Search**: Uses Firecrawl's search endpoint to find relevant Quora URLs based on your search criteria. - **💡 Intelligent Extraction**: Leverages Firecrawl's new Extract endpoint to pull user information from Quora profiles. - **⚙️ Automated Processing**: Formats extracted user information into a clean, structured format. - **💻 Google Sheets Integration**: Automatically creates and populates Google Sheets with lead information. - **✍️ Customizable Criteria**: Allows you to define specific search parameters to find your ideal leads for your niche. ### Tech stack: - Python ### Why it matters: The AI Lead Generation Agent automates the process of finding and qualifying potential leads from Quora, making it easier for sales and marketing teams to focus on building relationships and closing deals. It uses Firecrawl's search and the new Extract endpoint to identify relevant user profiles, extract valuable information, and organize it into a structured format in Google Sheets. This agent helps save hours of manual research and ensures that your sales team can efficiently build targeted lead lists.

Overview: Inimesed is an Android app that allows users to search their contacts by voice, utilizing the Android port of Pocketsphinx with Estonian acoustic models. Key features: - Voice-based contact search - No internet connection required - Estonian acoustic models Tech stack: Java Why it matters: Inimesed simplifies contact searching by voice, making it accessible even without an internet connection, and is specifically designed for the Estonian language, ensuring optimal performance and accuracy. Repository: Kaljurand/Inimesed GitHub stars: 61 GitHub description: An Android app that lets you search your contacts by voice. Internet not required. Uses Estonian acoustic models.

### Overview: ML Conformer Generator is a tool designed for generating 3D molecular conformations that are chemically valid and spatially similar to a reference shape. It utilizes Equivariant Diffusion Models (EDM) and Graph Convolutional Networks (GCN) to achieve this. ### Key Features: 1. **Shape-Guided Molecular Generation**: Generate novel molecules that conform to arbitrary 3D shapes, such as protein binding pockets or custom-defined spatial regions. 2. **Objective-Guided Generation**: Use reinforcement learning (RL) to steer molecular generation toward higher-scoring candidates, with support for custom scoring functions. 3. **Reference-Based Conformer Similarity**: Create molecules conformations that closely resemble a reference structure, supporting scaffold-hopping and ligand-based design workflows. 4. **Fragment-Based Inpainting**: Fix specific substructures or fragments within a molecule and complete or grow the rest in a geometrically consistent manner. 5. **Inertial Fragment Matching**: Generate molecules fragment-by-fragment by leveraging the physical properties of the shape descriptor, improving both shape similarity and chemical validity. ### Tech Stack: - JavaScript ### Why it Matters: ML Conformer Generator is a powerful tool for chemists and researchers looking to generate molecules with high spatial similarity and chemical validity. It simplifies complex molecular design tasks and can be used in various applications, including drug discovery and material science. ### Repository: Membrizard/ml_conformer_generator ### GitHub Stars: 61 ### GitHub Description: A tool for spatially-aware molecule design and optimisation via Equiv

### Overview: 2ByteASR is a powerful subtitle editing and translation tool designed for developers. It allows you to transcribe, proofread, translate, cut, and export subtitles in one convenient interface. With advanced ASR models, it supports bilingual subtitles and customizable glossaries, making it ideal for professional subtitle studios. ### Key Features: - **Subtitle Editing & Translation Workbench**: Transcribe, proofread, translate, cut, and export subtitles in one window. - **Supports Multiple Languages**: Transcribes English, Chinese, and Korean. - **Cloud and Local Capabilities**: Runs on your own computer or connects to a cloud service. - **Custom Glossaries**: Enhances subtitle quality with custom glossaries. - **Bilingual Subtitles**: Supports both Chinese and English subtitles. - **Real-Time Preview**: See the video and subtitles in real-time preview mode. - **Easy Setup**: Supports Windows and macOS, with a Mac version still experimental. ### Tech Stack: - Python ### Why It Matters: 2ByteASR simplifies the subtitle editing process, making it accessible and efficient for developers. Its advanced ASR models and bilingual capabilities make it a valuable tool for professional subtitle studios, allowing for seamless integration of speech and video, and easy customization with glossaries. ### Repository: Repository: **NOTBAD314/2ByteASR** GitHub Stars: 50 GitHub Description: A smarter subtitle studio powered by advanced ASR models. Refine transcripts with AI and custom glossaries, build bilingual subtitles, and edit speech and video together. ### Download:

Overview: LeapAlign is a post-training method that addresses the challenge of aligning flow matching models with human preferences, enabling gradient propagation to any generation step by constructing a two-step leap trajectory. Key features: - Leverages two-step leap trajectories for gradient propagation - Supports any generation step through backpropagation - Addresses the limitations of backpropagation through the full sampling trajectory Tech stack: Python Why it matters: LeapAlign simplifies the optimization process by enabling gradient flow to any generation step, making it particularly useful in scenarios where the full sampling trajectory is impractical or memory-intensive. This method enhances the efficiency and effectiveness of flow matching models in various applications, such as image generation and style transfer.

Overview: Child-Tuning is a source code repository for an EMNLP'21 paper that explores fine-tuning large language models to raise children. The project aims to develop effective and generalizable fine-tuning strategies for language models. Key features: - Fine-tuning large language models for child development - Effective and generalizable fine-tuning strategies - Source code for the EMNLP'21 paper Tech stack: Python Why it matters: This project provides a practical approach to fine-tuning language models for educational applications, offering insights into effective and generalizable fine-tuning strategies for language models. Repository: RunxinXu/ChildTuning Primary language: Python GitHub stars: 61

Overview: This project implements generative models such as Generative Adversarial Networks (GAN), Variational Autoencoders (VAE), Diffusion Models, and Autoregressive Models using PyTorch, PyTorch Lightning, and Hydra. Key features: - Implementations of GANs - Use of VAEs for variational inference - Diffusion models for generating images from noise - Autoregressive models for sequence generation Tech stack: PyTorch, PyTorch Lightning, Hydra Why it matters: This project showcases advanced techniques in generative modeling, making it valuable for researchers and developers interested in creating realistic images and sequences.

Overview: MemRoPE is an innovative method for training-free infinite video generation, addressing the issue of identity drift and quality degradation in autoregressive video diffusion models. By compressing evicted frames into Memory Tokens and applying Online RoPE Indexing, MemRoPE ensures temporal aggregation remains mathematically valid and maintains fidelity over time. Key features: - Dual-rate EMA for compressing past frames into Memory Tokens. - Online RoPE Indexing for storing keys without RoPE and applying dynamic position encoding. - Training-free approach, requiring no fine-tuning. Tech stack: Python, PyTorch, CUDA, Flash-Attn. Why it matters: MemRoPE enables up to 1-hour video generation with a fixed-size cache, resolving the identity drift and quality degradation issues commonly encountered in autoregressive video diffusion models. This method provides a significant improvement in video generation capabilities without the need for extensive training, making it highly valuable for applications requiring high-quality video synthesis.

Overview: Veo 3.1 Lite prompts offer tested image prompt recipes, including example renders, aspect ratios, model IDs, and cost per image. Key features: - 4 tested recipes with example renders, aspect ratios, model IDs, and cost per image. - One OpenAI-compatible endpoint at `https://api.apimart.ai/v1`, no monthly plan required. - Effective price of $0.064 per second, estimated cost of the whole set: $0.256. - Async by design, submits, takes `task_id`, polls `GET /v1/tasks/{id}`. Tech stack: Python. Why it matters: Teams can quickly switch between different models and pricing plans without setting up a subscription, ensuring cost transparency and flexibility. Repository: abkktyaqnmkmi/veo-3-1-lite-prompts Primary language: Python GitHub stars: 39 GitHub description: Veo 3.1 Lite prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`veo-3.1-lite`) and the cost per image. README excerpt: # Veo 3.1 Lite Prompts — 8 Tested Recipes With Outputs <!-- conv-kit:v1 --> <p align="center"> <img src="assets/badges/price.svg" alt="observed unit price"> <img src="assets/badges/billing.svg" alt="billing model"> <img src="assets/badges/compat.svg" alt="OpenAI-compatible endpoint"> </

Overview: This is an unofficial Rust-based client library designed to interact with the Google Gemini generative AI API. Key features: - Public Gemini endpoint - Vertex AI endpoint - Basic error handling - Support for custom models Tech stack: Rust Why it matters: The project aims to provide a Rust-based alternative to the Google AI Python SDK, offering a more efficient and potentially more secure way to interact with the Gemini generative AI API.

Overview: This guide provides an easy introduction to stable diffusion models, a type of neural network used for generating images from text descriptions. Key features: - Introduction to stable diffusion models - Explanation of the process of text-to-image synthesis - Instructions on setting up a Jupyter Notebook environment - Examples of running the tutorial on Google Colab or local Jupyter server Tech stack: Jupyter Notebook Why it matters: This tutorial serves as a comprehensive guide for developers looking to understand and implement stable diffusion models in their projects.