Projects

### Overview: DiffComplete is a novel diffusion-based approach designed to enable multimodal, realistic, and high-fidelity 3D shape completion. This project was presented at the NeurIPS 2023 conference and is built on the 3D-EPN dataset, which includes both training and evaluation data for various object classes. ### Key features: - **Diffusion-based 3D Shape Completion**: Utilizes diffusion models to generate high-quality 3D shapes from incomplete or noisy data. - **Realistic and High-Fidelity Completion**: Achieves realistic results by incorporating multimodal information from different sources. - **Versatile Object Classes**: Supports a wide range of object classes, including furniture, vehicles, and more. ### Tech stack: - **Programming Languages**: Python - **Data Structures**: Organized in a structured directory with subdirectories for training data and configuration files. ### Why it matters: - **Innovative Approach**: DiffComplete offers a novel solution for 3D shape completion, providing high fidelity and realism. - **Versatility**: The project supports a variety of object classes, making it applicable across different domains. - **Research Contribution**: Presented at NeurIPS 2023, contributing to the field of 3D shape completion and generative models. ### GitHub Stars: - **GitHub Stars**: 134 ### Repository: - **Repository**: JIA-Lab-research/DiffComplete - **Primary Language**: Python - **GitHub Description**: Official Codebase of "DiffComplete: Diffusion-based Generative

### Overview: PepGLAD is a state-of-the-art method for designing full-atom peptides with geometric latent diffusion. This project was presented at NeurIPS 2024 and is designed to address the challenges in peptide design by leveraging advanced geometric latent diffusion techniques. ### Key Features: - **Full-Atom Peptide Design**: Peptide sequences and structures are designed simultaneously using a geometric latent diffusion model. - **Geometric Latent Diffusion**: Utilizes geometric latent diffusion to capture the complex interactions between amino acids and the peptide backbone. - **Geometric Embedding**: Embeds geometric information into the latent space to improve the model's ability to handle complex peptide structures. ### Tech Stack: The project is implemented using Python, leveraging PyTorch for the neural network framework and OpenMM for molecular dynamics simulations. ### Why it Matters: PepGLAD offers a novel approach to peptide design that significantly improves the accuracy and efficiency of the design process, making it a valuable tool for researchers and developers in the field of computational chemistry and drug discovery.

Overview: FISOR is an innovative project that introduces a new approach to offline reinforcement learning, specifically focusing on the safety of reinforcement learning (RL) in real-world applications. This method leverages a feasibility-guided diffusion model to address hard constraints within the offline RL setting, aiming to enhance both performance and stability. Key features: - Implementation of FISOR for various RL environments, including Point Robot, Safety-Gymnasium, and Bullet-Safety-Gym. - Data quantity experiment and imitation learning experiment for MetaDrive. - Visualization of feasible regions for the Point Robot environment. Tech stack: Python. Why it matters: FISOR represents a pioneering effort in considering hard constraints within the safe offline RL setting, making it a significant contribution to the field of reinforcement learning. Its real-world applications, such as safety-critical scenarios generation and data center cooling system optimization, highlight its practical significance. Repository: ZhengYinan-AIR/FISOR GitHub stars: 134 GitHub description: [ICLR 2024] The official implementation of "Safe Offline Reinforcement Learning with Feasibility-Guided Diffusion Model".

### Overview: App-Controller is a framework that allows users to interact with applications using natural language commands. It simplifies the process of integrating and orchestrating APIs from various applications, making it easier for developers to add intelligent features to their apps. ### Key Features: - **Natural Language Command Integration**: Users can control applications by simply inputting natural language commands. - **API Orchestration**: Automatically identifies and orchestrates the best API calls to fulfill user instructions. - **Easy Integration**: Developers can register their application's API directory and let App-Controller handle the rest. - **Asynchronous and Concurrent Processing**: Enhances support for concurrent and asynchronous processing. ### Tech Stack: - **Python**: The primary language used for development. ### Why It Matters: App-Controller simplifies the interaction between users and applications, making the services or content more accessible and user-friendly. It reduces the need for complex interfaces and command sets, enhancing the overall usability and efficiency of the application.

### Overview: Grok Imagine API Guide: Image Generation and Editing on the Grok-Imagine Routes, Per-Image Pricing, Prompt Recipes with Example Renders and Async Task Polling Examples ### Key Features: - Per-image pricing - Prompt recipes with example renders and async task polling examples - One OpenAI-compatible endpoint at `https://api.apimart.ai/v1` - No monthly plan required - $0.015 per image, includes edits - One key, entire catalog - $1 minimum, pay as you go - Charge comes back in the response - Async by design, submit, take `task_id`, poll `GET /v1/tasks/{id}` ### Tech Stack: Python ### Why It Matters: - **One key, entire catalog.** The same `https://api.apimart.ai/v1` base URL and `Authorization` header reach Grok Imagine 1.5 (`grok-imagine-1.5-apimart`) and 300+ other image, video and language models — switch the `model` field, not your client. - **$1 minimum, pay as you go.** No subscription and no prepaid plan to size up front: top up from $1 and spend it on calls. There is no free quota to burn through first, so the price in this table is the price you pay. - **The charge comes back in the response.** Every call reports the amount billed (`cost` / `credits_cost`), so a spend number is read per

Overview: LetMeDoIt AI, version 3.0, is an advanced AI assistant that leverages the capabilities of ChatGPT API, Gemini Pro, AutoGen, and open-source LLMs. It enables seamless integration of multiple tools for both conversational engagements and local computing tasks. Key features: - Automates prompt engineering - Refines tool instruction refinement - Executes computing tasks - Creates action plans - Assembles agent creation - Selects multiple tools - Executes multiple steps - Implements Quality Control - Generates reports Tech stack: Python Why it matters: LetMeDoIt AI simplifies complex tasks by automating the process, making it more efficient and effective for developers and businesses alike. Repository: eliranwong/letmedoit GitHub stars: 134

### Overview: A live tracker that queries GitHub Search API for recently created and updated AI coding-agent skill repositories, ranks them by freshness and momentum rather than all-time stars, and publishes a self-updating README leaderboard via GitHub Actions. ### Key features: - Auto-updated every 15 minutes - Tracks trending Claude skills and AI agent repos - Ranks repositories by freshness and momentum - Refreshed data is current and reliable ### Tech stack: Python ### Why it matters: Fresh data every 15 minutes, ensuring that the list reflects the latest trends and changes in the AI coding-agent skill landscape.

### Overview: A local web search model for large language models (LLMs) that ranks results on your machine without relying on external APIs. This primitive RAG-like web search model context protocol (MCP) server runs locally and is designed for efficiency and privacy. ### Key features: - **No API keys.** Searches go through [ddgs](https://github.com/deedy5/ddgs), which asks several engines at once and skips blocked ones. - **Ranked locally.** A small MobileBERT model on plain numpy re-ranks results, fetching only the best pages. - **Clean text.** Pages are fetched with browser impersonation, keeping only the main content: no menus, cookie banners, or footers. Code blocks and tables keep their lines. PDFs are read too (text layer, first 50 pages), and docs sites that publish markdown copies of their pages are read from those. The same page under two URLs comes back once. - **Passages instead of pages.** `search` returns just the paragraphs that answer the question, about a tenth of the tokens, and `deep_research` can do the same per search term. When nothing matches well, `search` can ask you whether you want the full pages instead. - **Reads the sites that block scrapers.** Stack Overflow and the other Stack Exchange sites, Wikipedia, GitHub, Hacker News, PyPI, npm, and crates.io go through their free APIs: questions with the accepted answer first, articles by section, READMEs and raw files, HN threads with their

Overview: Azula is a Python package that implements diffusion models in PyTorch, providing a unified interface for various formalisms of generative diffusion models. Key features: - Unified formalism for diffusion models - Noise schedule, denoiser, and sampler components - Installation via pip or from the repository - Example code for training and sampling Tech stack: Python, PyTorch Why it matters: Azula simplifies the implementation of diffusion models by providing a single, convenient API, making it easier to experiment with and compare different models and techniques.

Overview: tenk is a tool that allows developers to ask questions about SEC filings and receive answers with direct citations, streamlining the process of understanding complex financial reports. Key features: - ✅ RAG over SEC filings with a local vector database - ✅ Auto-downloads filings from SEC EDGAR - ✅ Citations with links to source documents - ✅ Stock data via Yahoo Finance - ✅ Web search for data not in filings - ✅ Excel generation for tables and financial models - ✅ Export answers to PDF, DOCX, or Excel - ✅ Conversation memory with auto-summarization Tech stack: Python Why it matters: tenk automates the process of fetching and analyzing SEC filings, making it easier for developers to understand complex financial reports without the need for manual reading or extensive research.

### Overview: This repository contains the official code repository for the book "LLM from Scratch" (by길벗, 2025), which is a guide on how to build and train a language model from scratch. The book focuses on understanding how large language models operate at a fundamental level and provides detailed explanations, diagrams, and examples to help you build your own model. ### Key Features: - Code for building and training a language model from scratch. - Includes code for training a large language model similar to GPT. - Contains code for fine-tuning a pre-trained model. - Includes code for generating text based on a pre-trained model. - Contains detailed documentation and examples for each chapter. ### Tech Stack: - Jupyter Notebook ### Why it Matters: This repository is essential for developers and researchers interested in understanding the inner workings of large language models and building their own models from scratch. It provides a comprehensive guide to the process of training and fine-tuning language models, making it a valuable resource for anyone looking to explore this field. ### GitHub Stars: 135 ### Repository: rickiepark/llm-from-scratch

Overview: Comet is a high-performance hybrid vector store written in Go, designed for developers who want to understand how vector databases work and build one themselves. It supports various indexing strategies and search modalities, including Flat, HNSW, IVF, PQ, and IVFPQ, with features like reciprocal rank fusion, metadata filtering, and multi-query operations. Key features: - Vector Storage: Flat, HNSW, IVF, PQ, IVFPQ - Search Modalities: Vector, Full-Text, Metadata Filtering - Fusion Strategies: Weighted Sum, Reciprocal Rank Fusion, Max/Min Score - Data Structures: HNSW Graphs, Roaring Bitmaps, Bit-Sliced Index, Product Quantization Codebooks - Other Capabilities: Quantization, Soft Deletes, Serialization, Thread-Safe, Autocut Tech stack: Go Why it matters: Comet is a comprehensive solution for vector database developers, providing a unified package for hybrid retrieval, including various indexing strategies and search modalities, making it suitable for both developers and researchers. Repository: 0xnyn/comet GitHub stars: 133 GitHub description: A Vector Store written in Go - Supports hybrid retrieval over BM25, Flat, HNSW, IVF, PQ and IVFPQ Index with Quantization, Metadata Filtering, Reranking, Reciprocal Rank Fusion, Soft Deletes, Index Rebuilds and much much more.

Overview: OpenJet is an offline coding agent designed to run the Qwen3.8-27B model on local GPUs, providing up to 1.9x faster performance compared to llama.cpp. It supports various coding tools and is compatible with RTX 30, 40, and 50-series cards with 24 GB or more. Key features: - Runs the Qwen3.8-27B model on local GPUs - Supports RTX 30, 40, and 50-series cards with 24 GB or more - Compatible with Claude Code, Codex, OpenCode, and OpenClaw - Provides up to 1.9x faster performance than llama.cpp Tech stack: Python Why it matters: OpenJet offers a significant performance boost for coding tasks, making it an ideal solution for developers looking to optimize their workflow on their own hardware. Repository: L-Forster/open-jet GitHub stars: 129 GitHub description: An offline coding agent for local models. One command sets up the best model your GPU can run, and on 24 GB+ RTX 30/40/50-series cards Qwen3.8-27B runs up to 1.9x faster than llama.cpp.

### Overview: Open-Pomelli is an open-source project that provides an alternative to Google's Pomelli, offering customizable brand DNA, on-brand campaign concepts, platform-specific creatives, and AI-generated short-form videos. It is self-hosted and runs on SQLite, making it accessible and cost-effective. ### Key Features: - **Brand DNA Extraction**: Paste a website URL, and Playwright scrapes the site, generating an editable brand DNA profile. - **Editable DNA Review**: Customize tone, personality, and messages, and select native color pickers for primary and secondary palettes. - **Campaign Generation**: Choose a goal and direction, then generate 4 distinct on-brand concepts. - **Platform-Specific Asset Generation**: One-click produces ready-to-post creatives across 8 formats, including Instagram, LinkedIn, Facebook, and more. - **In-Browser Canvas Editor**: Every generated asset opens an editor with a 9-position grid for headline, body, and CTA blocks, and controls for size, color, background, font, and alignment. - **AI Photo Studio**: 30 product photography presets powered by `nano-banana-2-edit`, including Studio white, marble clean, urban street, golden hour, restaurant plated, and more. - **Animate**: Turn any asset, photoshoot, or fresh upload into a 3-12 second video via `seedance-lite-i2v`, with 480p, 720p, and 1080p options. ### Tech Stack: - TypeScript - SQLite

### Overview: SNDR Core Engine is a vLLM patch-overlay that optimizes Qwen3.6 and Gemma4 models for consumer NVIDIA GPUs, providing up to 1.5x faster inference and a 280K-token context. It is designed for frontiers of open-source LLMs and is family-tuned for Qwen3.6 and Gemma4. ### Key Features: - **Quantized Tool Calling**: Optimized for quantized models. - **MTP Speculative Decoding**: Speculative decoding for better performance. - **TurboQuant KV**: Optimized KV-cache quantization. - **Hybrid GDN**: Hybrid Generalized Denoising Autoencoder. - **26B/31B AWQ**: Advanced WaveNet Quality. - **256K Context**: High context capacity. - **321 Patches**: 321 patches covering various optimizations. ### Tech Stack: Python ### Why it Matters: SNDR Core Engine optimizes open-source LLMs for consumer NVIDIA GPUs, providing significant performance improvements and context capacity, making it a valuable tool for developers and researchers alike. ### Repository: Sandermage/sndr_core_engine ### Primary Language: Python ### GitHub Stars: 133 ### How it Works: SNDR Core Engine is installed in memory at boot, optimizing the vLLM engine for frontiers of open-source LLMs on consumer NVIDIA GPUs. ### Headline Numbers: - **1.5x Faster Inference** - **280

Overview: AdaFlash is an efficient speculative decoding framework that leverages on-policy distilled diffusion drafters and an adaptive length head. It is designed for high-performance inference and training, providing a reproducible benchmark for various models. Key features: - Asynchronous on-policy inference and training with hot-swappable draft checkpoints - On-Policy Distillation of diffusion drafters - Adaptive candidate-length initialization and training - SGLang-based serving and speculative decoding - Reproducible benchmarks for Vanilla, EAGLE-3, DFlash, MTP, OSD, and AdaFlash - Configurations and recipes for Qwen3, Qwen3-Coder, and Qwen3.5 model families - Training framework modified from [SpecForge](https://github.com/sgl-project/SpecForge) Tech stack: Python Why it matters: AdaFlash is a significant advancement in speculative decoding, offering improved performance and scalability compared to traditional methods. Its ability to adaptively control the length of candidate sequences makes it highly effective for real-time applications and large-scale inference tasks. Repository: ZinYY/AdaFlash GitHub stars: 133 GitHub description: [COLMW'26 Spotlight] AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters

Overview: This repository contains a multimodal chat interface powered by advanced AI models and tools, designed to enhance communication and information retrieval. Key features: - Web Interaction: Performs web searches and browses websites. - Wikipedia Tools: Finds relevant Wikipedia pages and articles. - Python Scripting: Runs Python scripts for computations and content management. - Content Management: Manages personal archives and sketchbooks. - Image Handling: Generates and searches for images. Tech stack: Python, Gradio, Amazon Bedrock, Anthropic Claude 3.5 Sonnet, Amazon Titan Text and Multimodal Embeddings, Amazon Titan Image Generator, OpenSearch, Amazon Polly, AWS Lambda. Why it matters: This tool enhances communication and information retrieval by providing a comprehensive interface for web browsing, knowledge retrieval, scripting, content management, and image processing, making it valuable for both personal and professional use.

### Overview: Docmancer is a powerful tool that securely restores and preserves the memory, rules, and instructions of your coding agents across multiple machines. It ensures that every agent has access to the same project context, making it easier for them to collaborate and learn from each other. ### Key Features: - **Secure Backup and Restore**: Docmancer securely backs up and restores your coding agents' sessions, memory, and setup, ensuring that the project context is preserved across different machines. - **Local First**: The tool operates entirely from local storage, ensuring that sensitive information is not exposed to the cloud. - **Project Mapping and Conflict Quarantine**: Docmancer helps in maintaining a consistent project context by mapping project files and quarantining divergent sessions, ensuring that every agent has access to the same project history. ### Tech Stack: - Python ### Why It Matters: Docmancer is essential for developers who work on multiple machines, ensuring that their coding agents have access to the same project context, making it easier for them to collaborate and learn from each other. It provides a secure and efficient way to manage and share project information across different environments, improving productivity and reducing errors.

Overview: Timething is a library designed to align text transcripts with audio recordings, providing precise time-codes for each word and character in the audio file. It uses powerful speech recognition models hosted by the Hugging Face AI community and is optimized for both GPUs and CPUs. Key features: - Aligns text transcripts with audio recordings - Outputs time-codes for each word and character in the audio file - Supports long form alignment for media content - Supports short form alignment for machine learning datasets - Aligns audio snippets with text transcriptions - Supports multiple audio formats including MP3, WAV, FLACC, and OGG/VORBIS Tech stack: Python with PyTorch and Hugging Face AI models Why it matters: Timething is essential for anyone working with audio-visual content, such as podcasters, audiobook producers, or researchers in the field of speech recognition. It provides a powerful tool for aligning audio recordings with transcripts, making it easier to verify and clean up audio datasets, and improving the accuracy of speech recognition models.

### Overview: InteractiveVideo is a user-centric framework designed for interactive video generation. It allows users to personalize videos with intuitive manipulation, enabling high-quality regional content control and precise motion control. ### Key Features: - **Personalize A Video**: Users can modify videos by adding or removing elements, changing colors, and adjusting the environment to create unique and engaging content. - **Fine-Grained Video Editing**: Users can make precise edits to videos, such as changing the background, adding or removing objects, and adjusting lighting to achieve the desired effect. - **Powerful Motion Control**: Users can control the motion of objects in videos, such as turning, shaking, or moving, to create dynamic and interactive scenes. ### Tech Stack: Python ### Why it Matters: InteractiveVideo is a valuable tool for content creators and developers looking to enhance their video editing capabilities. It allows for personalized and precise control over video content, making it an essential tool for anyone involved in video production or design.