Projects

### Overview: A no-code desktop app for generating high-quality synthetic datasets to fine-tune LLMs. ### Key features: - Topic planning and example generation - Quality validation via LLM Judge - Deduplication - HuggingFace Hub upload ### Tech stack: Python ### Why it matters: The tool automates the full dataset generation pipeline, making it easy to generate high-quality datasets without writing custom scripts.

Overview: DocRAG is a document-based Q&A chatbot application designed for interactive Q&A with your document library, leveraging state-of-the-art language models and vector databases for accurate and context-aware answers. Key features: - Multi-Format Document Upload - Intelligent Processing - Flexible Vector Storage - Conversational Interface - Source Tracking - Real-time Responses - Choice of LLMs - User Management & Authentication - Chat History & Insights - Simple Demo UI Tech stack: React frontend, Python/FastAPI backend, Node.js/Express backend, Streamlit demo application. Why it matters: DocRAG simplifies the process of interacting with your document library by providing a seamless, intelligent Q&A experience, ensuring accurate and context-aware answers through robust language models and efficient vector storage.

Overview: Nano-Banana is a high-speed, 4K text-to-image generator utilizing the open-source Gemini 3.1 Flash Image model for 5-character consistency. Key features: - Text-to-image generation - 4K resolution - 5-character consistency Tech stack: TypeScript Why it matters: Nano-Banana offers a fast and high-quality text-to-image solution, making it ideal for applications requiring precise and consistent image generation.

Overview: AssistAI is a macOS application designed to enhance productivity by indexing and interacting with local documents through a chat interface. Key features: - File Indexing - Chat Interface - Settings Customization Tech stack: Swift Why it matters: AssistAI leverages the Retrieval-Augmented Generation (RAG) model to provide context-aware responses, making it a valuable tool for enhancing productivity and document management. Repository: CocoaPriest/AssistAI GitHub stars: 7 GitHub description: macOS app to chat with your local documents

Overview: RETECO is a SemEval-2027 shared task focused on Reasoning-Oriented Retrieval, which involves retrieving documents based on the context and reasoning about when evidence applies and what has been established in a conversation. Key features: - Retrieval that requires reasoning about the temporal context and the established facts in a conversation. - Training and development data available on Hugging Face. - Starter kit includes BM25 baselines, official scorer, submission format checker, and a deterministic build script. Tech stack: Python. Why it matters: RETECO is a significant task in the field of information retrieval, focusing on how well a system can reason about the temporal and contextual information in a conversation to retrieve relevant documents. This task is important for improving the ability of information retrieval systems to understand and reason about the context of the queries they receive. Repository: DataScienceUIBK/RETECO Primary language: Python GitHub stars: 7 GitHub description: RETECO: A SemEval-2027 Shared Task on Reasoning-Oriented Retrieval

**Overview:** Nano-CUDA-MOE is a cutting-edge implementation of large-scale Mixture-of-Experts models optimized for high-performance computing. This project leverages the power of NVIDIA's RTX 3090 GPU and 64 GB of RAM to run Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash, achieving impressive throughput and efficiency. **Key features:** - **125B Model, 6B Active Experts**: Qwen3.8-Flash-Next runs on a 24 GB RTX 3090 with 6B active experts, achieving up to 97 tokens per second. - **DeepSeek-V4.1-Flash**: This model, with 359 GB of experts, utilizes NVMe for efficient memory management, decoding at approximately 8-12 tokens per second (~20 on repetitive output). - **CUDA-Optimized Architecture**: The project architecture ensures that the GPU holds the dense weights and a cache of the most used experts, while the CPU handles the missing experts, leading to high efficiency and low latency. - **Speculative Decoding with MTP Layer**: The model drafts up to 3 tokens, with rejection sampling on the GPU, maintaining the output distribution. - **Long Context and Sparse Attention**: The model processes long prompts efficiently, with a sparse attention mechanism and an int8 KV cache, capable of handling up to 256K tokens in 165 seconds (~1,550 tokens/s). **Tech Stack

Overview: Pulsing is a distributed actor framework built in Rust, designed for Python applications, providing location-transparent actor calls, streaming messages, and built-in cluster discovery without the need for Redis or etcd. Key features: - Location-transparent actor calls - Streaming messages - Built-in cluster discovery Tech stack: Rust Why it matters: Pulsing simplifies the development of distributed systems by providing a communication backbone for Python applications, making it easier to build and manage distributed systems without the complexity of setting up and managing a cluster. Repository: DeepLink-org/Pulsing GitHub stars: 7 GitHub description: Pulsing is a distributed actor framework that provides a communication backbone for building distributed systems, with specialized support for AI applications. README excerpt:

### Overview: Karma AI: The Perfected Garud Puran Guide is a professional, high-performance spiritual AI assistant powered by Gemini 2.0 Flash, PostgreSQL, and Streamlit. It analyzes deeds based on the ancient wisdom of the Garud Puran, offering guidance on Karma, Swarg (Heaven), and Nark (Hell). ### Key Features: - **🚀 Gemini 2.0 Flash**: Ultra-fast, intelligent spiritual insights. - **📚 Agentic RAG**: Uses ChromaDB to ensure every answer is grounded in high-quality Garud Puran datasets. - **🎙️ Voice-First Interaction**: Voice Input: Speak your questions naturally. Voice Output: Listen to AI-generated spiritual summaries. - **🏛️ Persistent Memory**: Every interaction is stored in PostgreSQL for your personal spiritual history. - **💎 Premium UI**: A sleek, dark-themed Streamlit interface designed for focus and peace. ### Tech Stack: - Python ### Why It Matters: Karma AI provides a personalized spiritual guide based on the Garud Puran, offering insights into Karma, Swarg, and Nark, helping users understand their deeds and spiritual path. ### Repository: - Edge-Explorer/Voice-Input-Karma_gemini - GitHub Stars: 7 - GitHub Description: Karma Ai where u can ask any query related to Garud Puran ### Project Structure: ```text Voice-Input-Karma_gemini/ ├── backend/ # FastAPI Logic & AI Core │ ├── main.py # Entry point (API) │ ├── core

Overview: This project, Negative-Space Learning (NSL), is a self-improving coding agent designed to free disk space by executing cleanup strategies on disposable Linux containers. It avoids traditional reward models, human labels, and model grading, focusing solely on measurable outcomes. Key features: - Self-improving agent - Uses disposable containers for cleanup - Measures actual disk space freed - Logs every attempt as training data Tech stack: Python Why it matters: This innovative approach challenges the limitations of traditional reinforcement learning methods by focusing on measurable outcomes rather than relying on fragile signals like rewards or human labels. The project demonstrates the potential for continuous improvement without the need for external validation or feedback mechanisms, making it a significant contribution to the field of artificial intelligence and machine learning.

Overview: llm-agent is a Java library designed to simplify the integration of Large Language Models (LLMs) into Java applications. Key features: - Easy integration with various LLMs - Support for multiple LLM providers - Configuration flexibility - Real-time updates and notifications Tech stack: Java Why it matters: llm-agent provides a straightforward and efficient way to incorporate LLMs into Java projects, making it easier for developers to build applications with AI capabilities.

### Overview: Eric-LLMs/llm-lab is a comprehensive deep-dive LLM engineering lab featuring a range of cutting-edge technologies and methodologies. It includes features such as Autonomous Agents (ReAct/AutoGPT), Fine-Tuning of Llama/Mistral models, Large Model Deployment, Advanced RAG, Function Calling, Frameworks (LangChain, Semantic Kernel, OpenAI), and daily tracking of the latest research in the LLM stack. The project is designed to provide a hands-on, practical approach to understanding and implementing state-of-the-art LLM engineering practices. ### Key features: - Autonomous Agents (ReAct/AutoGPT) - Fine-Tuning Llama/Mistral (SFT/DPO) - Large Model Deployment (DeepSeek 671B / 2.5-bit) - Advanced RAG (Hybrid Search) - Function Calling (Stream/Text-to-SQL/External APIs) - Frameworks (LangChain, Semantic Kernel, OpenAI) - Daily SOTA Paper Tracking ### Tech stack: Jupyter Notebook ### Why it matters: Eric-LLMs/llm-lab offers a practical, hands-on approach to LLM engineering, covering everything from theoretical architecture to practical implementation. It provides a comprehensive guide to creating autonomous agents, fine-tuning models, deploying large models, and more, making it a valuable resource for developers and researchers in the field of large language models.

Overview: localpal is a Windows command-line application that allows developers to run local language models in Delphi Object Pascal, providing in-process inference, model catalog downloads, persistent chat sessions, custom personas, and hardware benchmarking. Key features: - In-process inference with llama.cpp bindings - Model catalog and downloads from Hugging Face - Persistent chat sessions and session management - Custom personas and session export/import - Benchmarking for real prompt-processing and token-generation speeds Tech stack: Delphi, llama.cpp bindings, SQLite Why it matters: localpal simplifies the integration of large language models into Delphi projects, offering a convenient and efficient way to run models locally without the need for servers or cloud services. Repository: FMXExpress/localpal GitHub stars: 7 GitHub description: Run local LLM models in Delphi Object Pascal with llama.cpp.

Overview: This project is a simple implementation of speculative model serving using Apache Flink, an extension of a basic model service implementation. It is based on a blog post by Lightbend. Key features: - Speculative model serving with Apache Flink - Includes client, data, model, protobufs, query, and server modules - Provides both Scala and Java implementations Tech stack: Scala Why it matters: This project showcases the speculative model serving feature of Apache Flink, demonstrating how it can be used in a practical application. Repository: FlinkML/flink-speculative-modelServer GitHub stars: 7 GitHub description: Speculative model serving with Flink

### Overview: Vision Studio-X is a professional AI-powered desktop application for local image and video generation. It offers a seamless experience without the need for cloud services, running entirely on your machine. The application supports FLUX.1, Stable Diffusion 3.5, Stable Diffusion XL, SD 1.5, and LTX Video, among others, using Electron, React, FastAPI, and PyTorch. ### Key Features: - **Image Generation**: Supports FLUX.1, Stable Diffusion 3.5, Stable Diffusion XL, and SD 1.5. - **Video Generation**: Supports LTX Video and AnimateDiff. - **Guided Edit Tools**: Provides background editing capabilities. ### Tech Stack: Electron, React, FastAPI, PyTorch. ### Why it Matters: Vision Studio-X is designed for a professional environment, offering local AI image and video generation without the need for cloud services. This makes it ideal for businesses and individuals who prefer to keep their data and processes under their control. The application's ability to run on NVIDIA CUDA, Apple Silicon, and CPU makes it versatile and compatible with a wide range of hardware setups. ### Repository: - **Repository**: Git-Rocky-Stack/vision-studio - **GitHub Stars**: 7 - **GitHub Description**: Free, open-source (MIT) desktop app for local AI image & video generation - FLUX.1, SD 3.5, SDXL, LTX Video, AnimateDiff. Electron + React + FastAPI/PyTorch

Overview: This project, `Goodnight77/qdrant-exporter`, is a Prometheus exporter designed to monitor and expose metrics from Qdrant, a vector database. It provides a simple way to observe the performance and health of Qdrant collections through Prometheus and Grafana. Key features: - Scrapes Qdrant API for collection data - Exposes exporter metrics at `/metrics` - Works with Prometheus and Grafana Tech stack: Go Why it matters: This exporter makes Qdrant more observable, allowing developers to monitor and debug their vector database more effectively.

Overview: This project fine-tunes large language models (LLMs) using BERT to answer questions based on the SQuAD 2.0 dataset. Key features: - Fine-tuning of pre-trained language models - Accurate question answering with high F1 and EM scores Tech stack: Jupyter Notebook Why it matters: This project demonstrates the effectiveness of fine-tuning LLMs for practical applications in natural language processing, specifically in answering questions from provided contexts.

### Overview: CleanPilot AI 是一款面向智能清洁设备的企业级多智能体服务平台，旨在通过先进的多智能体技术提升设备运营效率和用户体验。该平台采用调度 Agent、知识问答、故障诊断、用户运营等三个功能Agent，结合混合 RAG检索、受控 Skill、分层语义记忆、JWT/RBAC与React SSE流式交互，实现了从知识运营到服务结果的完整业务闭环。 ### Key features: - **混合 RAG检索**：通过Chroma向量召回、FTS5 BM25关键词召回、RRF融合和模型精排提升检索质量。 - **知识可靠性**：使用Chunk级SHA-256、稳定向量ID、状态跟踪和失败重试替代文件级粗粒度去重，支持单片段修复和文件级重建。 - **可交付能力**：实现JWT/RBAC、账户数据隔离、分层语义记忆、管理员知识运营和SSE逐Token输出，覆盖关键权限、检索、记忆及流式协议。 - **业务闭环**：调度Agent自动路由三个功能Agent，故障诊断和用户运营报告由受控Skill执行，工具参数由可信运行时上下文强制绑定，阻止跨账户查询和越权调用。 ### Tech stack: - Python ### Why it matters: CleanPilot AI通过先进的多智能体技术，显著提升了智能清洁设备的运营效率和用户体验，为企业级智能清洁设备的市场推广和客户服务提供了强有力的支持。其强大的检索能力和知识

Overview: This repository contains the official code for the paper "Test-Time Compute Games" by Ander Artola Velasco, Dimitrios Rontogiannis, Stratis Tsirtsis, and Manuel Gomez-Rodriguez, published in 2026. It models the LLM-as-a-service market as a game where providers compete by selecting test-time compute levels to maximize profit, introducing a Reverse Second-Price Auction mechanism to align incentives with social welfare. Key features: - The repository is organized into different directories for configuration files, notebooks, scripts, and source code. - It includes benchmark datasets for GSM8K, AIME, and GPQA, and a package source code for various models and compute budgets. - The README provides instructions on how to clone the repository, create a virtual environment, and install dependencies. - The paper's main findings are analyzed through Jupyter notebooks for GSM8K, AIME, and GPQA, generating plots and tables that are saved in the `figures/` directory. - The repository structure is detailed, and the primary language is Jupyter Notebook. Tech stack: Jupyter Notebook Why it matters: This project provides a comprehensive solution for understanding and analyzing the LLM-as-a-service market, offering insights into provider competition and the social welfare alignment mechanism introduced in the paper. GitHub stars: 7 Repository: Human-Centric-Machine-Learning/strategic-ttc

### Overview: 这是一套针对图片、视频、音频、数字人和多模态创作工作台的“证据优先”AIGC产品拆解技能。它通过页面截图、聊天记录、画布、资产卡、任务状态、模型选项和错误提示等多级证据，逐层还原用户旅程、Agent契约、工具调用与运行结果、项目全局上下文、模型接入与路由边界等关键信息，提供详尽的HTML产品架构报告，帮助开发者理解并优化AIGC产品的设计和实现。 ### Key features: 1. **用户旅程**：从输入需求到最终结果的真实旅程还原。 2. **Agent契约**：整理六类输入、五类输出、可观察判断、功能性工具、全局上下文读写、完成条件和异常边界。 3. **功能等价System Prompt**：为单个Agent编写可直接使用的功能等价Prompt，复现页面行为，但不冒充产品官方原文，也不声称读取隐藏思维链。 4. **产品全景架构**：合并用户交互流、Agent控制流、工具调用流、上下文流和媒体资产流，补充数据实体、ER图、端到端时序、模型路由、As-Is/To-Be和关键风险。 5. **HTML可视化报告**：用泳道、节点链、分层图、状态图、风险矩阵和证据追溯表生成可交付页面。 ### Tech stack: Python ### Why it matters: 这套方法

Overview: Strata-GLM is a GLM-5.3-Flash engine optimized for gaming hardware, designed for high throughput and low latency. It leverages a VRAM-resident expert system with flat routing and chunked MMQ prefill, achieving impressive performance on a 24GB GPU and 90GB of RAM. Key features: - High throughput decoding and prefill capabilities - Flat routing for expert residency - Chunked MMQ prefill for efficient memory usage - 512K context length - 45 layers with mixed hyper-connections and KDA attention - 45 layers × 288 routed experts - 320B total model size, 18B active Tech stack: C++ Why it matters: Strata-GLM is optimized for gaming hardware, providing high throughput and low latency decoding and prefill capabilities. Its design allows for efficient memory usage and expert residency, making it suitable for real-time applications and high-performance computing environments.