Projects

Overview: DegreeGuru is a project designed to create a custom AI RAG chatbot using Langchain.js, Upstash Vector, and OpenAI. It includes a built-in crawler for scraping data from websites and is fully domain agnostic. Key features: - 🕷️ Built-in crawler for scraping data from any website - ⚡ Fast answers using Upstash Vector and real-time data streaming - 🛡️ Includes rate limiting to prevent API abuse Tech stack: TypeScript, Scrapy, Langchain.js, Upstash Vector, OpenAI, Vercel AI, Upstash Why it matters: This project is ideal for anyone looking to create their own AI RAG chatbot with custom data, making it a valuable resource for educational and research purposes.

### Overview: SIM-CoT is an official implementation of "SIM-CoT: Supervised Implicit Chain-of-Thought" presented at ICLR 2026. It adds step-level supervision to implicit reasoning methods like Coconut and CODI, enabling more compact reasoning while still maintaining high accuracy. ### Key Features: 1. **Step-Level Supervision**: An auxiliary decoder learns to recover each reasoning step from its corresponding latent during training. 2. **Compact Reasoning**: Remove the decoder at inference to keep reasoning compact; reattach it for optional step inspection. 3. **Paper Case Visualization**: Based on Figure 4 and Appendix I.1, the animation illustrates latent reasoning and optional step decoding; it is not a live model run. ### Tech Stack: Python ### Why It Matters: SIM-CoT addresses the instability in implicit reasoning methods by adding step-level supervision, making the model more compact and improving accuracy without sacrificing performance.

Overview: This project is an HTTP server designed to test the Coqui STT project, a successor to Mozilla DeepSpeech. It allows developers to evaluate speech-to-text models without the need for training. Key features: - HTTP server for testing Coqui STT models - Supports Coqui STT 1.0 APIs - Available on PyPI and from sources - Configured with a YAML file for server settings Tech stack: Python Why it matters: This server is crucial for developers who want to evaluate and test speech-to-text models without the need for training, making it an essential tool for developers working on natural language processing projects.

### Overview: MooER: Moore-threads Open Omni model for speech-to-speech interaction is a comprehensive toolkit designed for speech-to-speech interaction tasks. It includes end-to-end models for speech translation, speech recognition, and more, with a focus on open-source and modular design. ### Key Features: - **Speech Translation**: Supports real-time speech translation between multiple languages. - **Speech Recognition**: Accurate speech recognition for various languages. - **Multi-Task Learning**: Integrates multiple tasks such as speech translation and recognition. - **Omni Model**: A unified model that can handle various tasks including speech translation and recognition. ### Tech Stack: Python ### Why It Matters: MooER provides a powerful and flexible framework for developers looking to integrate speech-to-speech interaction capabilities into their projects. Its modular design and open-source nature make it an excellent choice for both research and commercial applications.

Overview: NeMo-Speech.cpp is a lightweight native C++ inference runtime designed for efficient speech model inference, providing day-0 support for the latest models from NVIDIA NeMo Speech. Key features: - Native C++ inference - Broad hardware support - Supports speech recognition, speaker diarization, text-to-speech, and more Tech stack: C++ Why it matters: NeMo-Speech.cpp is NVIDIA's official solution for local speech inference, offering blazing fast performance and seamless integration with the latest models from NVIDIA NeMo Speech.

### Overview: Glif is a versatile creative agent that allows users to describe what they want to create, and it automatically selects the appropriate models and performs the necessary tasks across images, video, and audio. This includes transcribing, rendering HTML, searching the web, running code, and chaining multi-step media jobs. The Glif MCP Server and Plugin integrates seamlessly with popular models like Claude, ChatGPT, Codex, Cursor, and more, providing a streamlined workflow for generating high-quality content. ### Key Features: - **Endpoint:** `https://glif.app/api/mcp` (Streamable HTTP, JSON-RPC 2.0) - **Auth:** OAuth 2.1 sign-in with your Glif account. Plain HTTP clients can send a `glif_v1_...` API key from [glif.app/settings/api-tokens](https://glif.app/settings/api-tokens) as a bearer token. - **Install Page:** https://glif.app/mcp - **Skills:** Available skills include getting started, web visuals, app icons and logos, and app promo materials. ### Tech Stack: - JavaScript ### Why It Matters: Glif offers a powerful and flexible solution for generating high-quality content across various media types, making it an essential tool for creatives and developers looking to streamline their workflow.

### Overview: Speech synthesis, speech recognition, voice activity detection, and speaker diarization for voice conversations. The library is built on ggml and supports Qwen3-TTS, Irodori-TTS, Qwen3-ASR, ReazonSpeech, and Silero VAD. It runs on Metal, Vulkan, and the CPU, and is compatible with an OpenAI-compatible server and live transcription. ### Key features: - **Speech Synthesis**: Converts text to speech. - **Speech Recognition**: Transcribes audio to text. - **Voice Activity Detection**: Identifies where someone is speaking in recordings. - **Speaker Diarization**: Separates conversations into individual speakers. ### Tech stack: C++ ### Why it matters: This library simplifies the process of integrating voice technology into applications, making it easier for developers to implement speech synthesis, recognition, and other voice-related functionalities without needing to dive into complex APIs or frameworks.

### Overview: MineEcho is a local-first Memory OS designed for personal AI assistants. It combines raw interaction memory, daily summaries, and skill routing to provide a durable and efficient assistant experience. ### Key features: - **Memory Tree + Knowledge Base + Skill Routing + Cost Layer:** Stores raw interaction memory, daily summaries, and semantic recall evidence. - **Wiki++ Knowledge System:** Manages imported sources as raw files, wiki pages, chunk indexes, and alignment candidates. - **Skill and AI App Convergence:** External AI apps are registered as callable skills and participate in the same trigger, routing, health, and workflow surface. - **TokenLess Cost Layer:** Compresses long command logs, tool outputs, URL imports, and document extraction results before entering memory or context. - **Local-first Release Posture:** Runtime data, provider keys, chat history, audio, and metrics live under local MineEcho state. ### Tech stack: TypeScript ### Why it matters: MineEcho offers a memory-first assistant loop, Wiki++ knowledge system, skill and AI app convergence, and token-less cost layer, making it a unique solution for personal AI assistants.

Overview: This repository contains evaluation code and dataset for the paper "Are We on the Right Way for Evaluating Large Vision-Language Models" by Lin Chen and Jinsong Li. It aims to address two primary issues in current evaluation works: visual content being unnecessary for many samples and unintentional data leakage in LLM and LVLM training, providing an elite vision-indispensable multi-modal benchmark. Key features: - Evaluation code for MMStar - Support for online Leaderboard - Curate online test set, MMStar-test - Display 6 core capabilities and 18 detailed axes Tech stack: Python Why it matters: This benchmark helps in accurately evaluating the performance of large vision-language models, ensuring fair and accurate assessments of their capabilities and limitations.

Overview: RodinHD is a high-fidelity 3D avatar generation project that leverages diffusion models for efficient and accurate avatar creation. This innovative approach aims to provide realistic and lifelike avatars for various applications. Key features: - High-fidelity 3D avatar generation - Utilizes diffusion models for efficient avatar creation - Supports multi-scale vae features and triplane fitting for improved performance Tech stack: Python Why it matters: RodinHD offers a significant advancement in the field of 3D avatar generation by combining diffusion models with triplane fitting techniques, resulting in high-fidelity and realistic avatars. This project is particularly useful for applications requiring high-quality, realistic representations of avatars, such as in virtual reality, gaming, and beyond.

### Overview: NiT: Native-resolution Image Synthesis NiT is a model that explicitly learns varying resolutions and aspect ratios within its denoising process, significantly improving training efficiency and generalization capability. It achieves state-of-the-art results on both $256\times256$ ($2.08$ FID) and $512\times512$ ($1.48$ FID) benchmarks in class-guided ImageNet generation. NiT can also generalize to arbitrary resolutions, such as $4.52$ FID on $1024\times1024$ resolution and $4.11$ FID on $432\times768$ resolution. ### Key Features: - Native-resolution image synthesis - Explicit learning of varying resolutions and aspect ratios - SOTA results on $256\times256$ and $512\times512$ benchmarks - Generalization to arbitrary resolutions ### Tech Stack: Python ### Why it Matters: NiT is the first work to explicitly model native resolution image synthesis, achieving state-of-the-art results on both $256\times256$ and $512\times512$ benchmarks in class-guided ImageNet generation. It can generalize to arbitrary resolutions, improving efficiency and performance. ### Repository: WZDTHU/NiT ### Primary Language: Python ### GitHub Stars: 218 ### GitHub Description: NeurIPS 20

Overview: OBELICS is an open, massive, and curated collection of interleaved image-text web documents containing 141M documents, 115B text tokens, and 353M images. It serves as a valuable resource for researchers and developers interested in web document analysis and natural language processing. Key features: - Download WARC files from Common Crawl dumps - Extract HTML files from WARC files - Simplify HTML DOM trees - Convert simplified DOM trees to another structure - Perform web document extraction - Perform web document filtering - Perform line deduplication - Visualize the results Tech stack: Python Why it matters: OBELICS provides a comprehensive dataset for analyzing web documents, enabling researchers to study the structure and content of web pages in detail. Its open nature and large size make it an essential resource for academic research and practical applications in NLP and web document analysis.

Overview: PlayHT Python SDK provides a powerful toolset for converting text into high-quality audio streams using AI text-to-speech technology, including streaming and non-streaming options. This SDK simplifies the integration of Play's conversational voice AI models into Python applications. Key features: - Stream text-to-speech in real-time, synchronous or asynchronous. - Generate non-streaming text-to-speech, synchronous or asynchronous. - Use Play's pre-built voices or your own custom voice clones. - Stream text from LLM, and generate audio stream in real-time. - Supports WAV, MP3, Mulaw, FLAC, and OGG audio formats as well as raw audio. - Supports 8KHz, 16KHz, 24KHz, 44.1KHz and 48KHz sample rates. Tech stack: Python, aiohttp, filelock, grpc, requests, websockets. Why it matters: This SDK enables developers to easily integrate Play's AI text-to-speech capabilities into their applications, providing a seamless and high-quality audio experience for users.

### Overview: Gemini CLI is a command-line interface designed for interacting with Google Gemini models, enabling developers to harness the capabilities of multimodal AI through a simple command-line interface. ### Key Features: - **Command-Line Interface (CLI) for Google Gemini Models**: Provides a straightforward way to interact with various generative models. - **System Commands**: Supports system commands for internal operations, such as selecting generative models and chat history management. - **Configuration File**: Stores generative model settings and chat history, allowing for customization and persistence. ### Tech Stack: Go ### Why It Matters: Gemini CLI simplifies the interaction with complex AI models, making it easier for developers to leverage the power of multimodal AI in their projects.

### Overview: This project automates the Google Gemini Notebook (formerly NotebookLM) by integrating it with the MCP (Model Context Protocol) and providing a local HTTP REST API. It includes citation-backed Q&A, full Studio generation, and supports multi-account rotation with auto-reauth across personal and Google Workspace accounts. ### Key Features: - Citation-backed Q&A system - Full Studio generation (audio, video, infographic, report, presentation, data table) - Multi-account rotation with auto-reauth - Built on a dual transport of the internal `batchexecute` RPC API and Playwright as a fallback - Node.js 22 compatible and updated to handle new DOM changes in v2.3.0 and v3.0.0 ### Tech Stack: TypeScript ### Why it Matters: This project is essential for developers looking to automate Google NotebookLM at scale, providing a reliable and efficient way to integrate with the service. It supports citation-backed Q&A, full Studio generation, and multi-account rotation, making it a valuable tool for automating Google's NotebookLM.

Overview: AxBench is a scalable benchmark designed to evaluate interpretability techniques on two key axes: concept detection and model steering. It provides a comprehensive framework for generating synthetic data and evaluating a variety of interpretability methods, including finetuning and prompting baselines. Key features: - Scalable evaluation harness - Comprehensive implementations of 10+ interpretability methods - 16K concept training data - Two pretrained SDL models - LLM-in-the-loop training Tech stack: Python Why it matters: AxBench offers a valuable resource for developers and researchers looking to benchmark and evaluate interpretability techniques, providing a clear and comprehensive way to measure the effectiveness of various methods in improving LLM interpretability.

### Overview: 从 MiniMind 源码读起，延伸到现代大模型技术体系的中文学习笔记。主线逐行精读预训练 / SFT / DPO / PPO / GRPO 与训练机制；附录 17 篇进阶卷覆盖量化、投机解码、RLHF 全景、模型代际史等 MiniMind 没涉及、但进阶绕不开的主题。 ### Key features: - 逐行对照 MiniMind 源码，讲解每段代码的功能和设计。 - 包括 MiniMind2 主线、MiniMind-3 / Qwen3-style 代码对比、真实实验证据、训练机制、版本演进、实验与延伸卷。 - 详细解释 MiniMind 源码中的技术细节，如 RMSNorm、归一化演进、RoPE、GQA、SwiGLU、MoE 等。 - 附录涵盖了量化、投机解码、RLHF 全景、模型代际史等进阶主题，帮助读者深入理解现代大模型技术体系。 ### Tech stack: Python ### Why it matters: 从 MiniMind 源码读起，可以深入了解现代大模型的训练机制和技术脉络，帮助读者从源码层面真正理解一个语言模型的全流程。

Overview: dictate.js is a small JavaScript library designed for browser-based real-time speech recognition, utilizing Recorderjs for audio capture and a WebSocket connection to the Kaldi GStreamer server for speech recognition. Key features: - Real-time speech recognition - WebSocket communication - Compatible with various browsers Tech stack: JavaScript Why it matters: dictate.js simplifies the process of integrating speech recognition into web applications, making it easier for developers to implement real-time dictation features without the need for complex server-side setups. Repository: Kaljurand/dictate.js Primary language: JavaScript GitHub stars: 217 GitHub description: A small JavaScript library for browser-based real-time speech recognition.

Overview: SimplyRetrieve is a lightweight and user-friendly GUI and API platform designed for the Retrieval-Centric Generation (RCG) approach, focusing on efficient and interpretable AI tool development. It features a private and customizable knowledge base, retrieval tuning module, and advanced prompt engineering capabilities. Key features: - 📚 GUI and API-based Retrieval-Centric Generation platform - 🔧 Retrieval Tuning Module with Prompts Engineering, Tool Configuration, and Retrieval Analysis - 🏡 Private Knowledge Base Constructor - 🧠 Fully localized, private, and lightweight access to various sizes of open-source Large Language Models (LLMs) - 🚥 Multi-user concurrent access to UI, utilizing Gradio's queue function Tech stack: Python Why it matters: SimplyRetrieve aims to contribute to the development of safe, interpretable, and responsible LLMs by providing an open-source tool for implementing the RCG approach. It enables the machine learning community to explore the use of LLMs in an efficient and local manner, while maintaining privacy and local implementation.

Overview: K2 is a large language model designed for geoscience knowledge understanding and utilization, developed as part of a paper accepted at WSDM-2024. It is an open-source model that combines pretraining on geoscience literature with fine-tuning on knowledge-intensive instruction data. Key features: - Open-source implementation - Geoscience-specific pretraining - Fine-tuning with GeoSignal data - Available on Hugging Face for further fine-tuning Tech stack: Python Why it matters: K2 provides a foundational model for geoscience language understanding and utilization, offering a significant improvement in performance compared to similar models on both objective and subjective tasks. Its open-source nature and ability to be fine-tuned on specific datasets make it a valuable tool in geoscience research and applications. Repository: davendw49/k2 GitHub stars: 217 GitHub description: Code and datasets for paper "K2: A Foundation Language Model for Geoscience Knowledge Understanding and Utilization" in WSDM-2024