Projects

### Overview: MediVision AI is a fully open-source chest X-ray diagnostic assistant that combines a ResNet-50 classifier, Grad-CAM explainability, and a RAG pipeline over clinical guidelines to draft structured reports. The project is built entirely on free, open-source components, including no paid APIs. ### Key features: - **ResNet-50 Classifier**: Trained on ImageNet to classify chest X-ray findings. - **Grad-CAM**: Highlights the driving region for better understanding of the image. - **RAG Pipeline**: Retrieves relevant clinical guidance from ChromaDB and BioBERT embeddings, then drafts a structured report using Ollama (Mistral-7B). ### Tech stack: Python ### Why it matters: MediVision AI is a valuable tool for healthcare professionals, providing a user-friendly interface to assist in the interpretation of chest X-rays. The use of open-source components ensures that the project is accessible and reliable, without the need for expensive APIs or paid services. ### Repository: ashishallu/medivision.AI ### GitHub stars: 7 ### Primary language: Python ### GitHub description: MediVision AI is a chest X-ray diagnostic assistant that combines a ResNet-50 classifier, Grad-CAM, and a RAG pipeline over clinical guidelines to draft structured reports. The project is built entirely on free, open-source components, including no paid APIs.

### Overview Luthor is a Retrieval-Augmented Generation (RAG) system designed for law firms to enable lawyers to efficiently "talk to their data". This system allows legal professionals to upload documents (e.g., memos) to a vector database stored in Pinecone and subsequently query this information using a chatbot-like interface built with Streamlit. ### Key Features - Document upload and processing (supports .txt, .pdf, and .docx files) - Text preprocessing and segmentation - Vector embedding generation using OpenAI's API - Efficient document storage and retrieval using Pinecone - Natural language querying interface - Context-aware answer generation ### Tech Stack Streamlit, OpenAI API, Pinecone, Docker, AWS ECS, ECR, Application Load Balancer, AWS Secrets Manager, AWS ECS Cluster, ECR Repository, Application Load Balancer, VPC, Security Groups, IAM Roles ### Why It Matters Luthor simplifies the process of integrating legal data into the legal workflow, enabling lawyers to quickly access relevant information and generate accurate answers, improving efficiency and reducing the time spent on repetitive tasks. ### Repository boemer00/luthor ### GitHub Stars 7 ### Primary Language Jupyter Notebook ### GitHub Description A RAG system designed for law firms to enable lawyers to efficiently "talk to their data".

**Overview:** OpenTallas is an open research project focused on accelerating model-specific inference for large language models. It aims to make this architectural path to AI systems more accessible and extensible, with a primary emphasis on improving silicon performance. **Key features:** - Open research towards model-specific mask-ROM LLM inference silicon. - Analytical studies, ABI 3.0 compiler/runtime, and cycle model. - Synthizable RTL and open-PDK physical evidence. - Gated against itself by a 19-rung board that is deliberately mostly red. **Tech stack:** Verilog. **Why it matters:** OpenTallas provides a credible path toward AI systems that can respond at interactive timescales, challenging the notion that model-specific inference is only a paper architecture. It connects model accounting, architecture simulation, and physical evidence in one evidence chain, making it independent of specific hardware implementations. **Repository:** bojieli/OpenTallas **GitHub stars:** 7 **GitHub description:** Open research toward model-specific mask-ROM LLM inference silicon: analytical studies, an ABI 3.0 compiler/runtime and cycle model, synthesizable RTL, and open-PDK physical evidence — gated against itself by a 19-rung board that is deliberately mostly red. **GitHub description:** More than 10,000 tokens per second per user is now a publicly reported silicon result. Taalas reports **16,960 tokens/s** per user for its fabricated HC1 accelerator running Llama 3.1 8B at batch one. This result

### Overview: KnowFlow is a powerful hybrid Retrieval-Augmented Generation (RAG) system designed for intelligent document processing and querying. Combining semantic search with knowledge graph reasoning, it offers advanced document processing, intelligent chunking, and secure document isolation, making it an essential tool for modern document management and querying. ### Key Features: - **Advanced Document Processing**: Multi-format support, configurable chunk size and overlap, parallel batch processing with S3 storage, document status tracking (PENDING, PROCESSING, INDEXED, FAILED). - **Hybrid RAG + Knowledge Graph Architecture**: Dense semantic embeddings via Google Gemini + pgvector, structured knowledge extraction to Neo4j, multi-hop reasoning through graph relationships, automatic entity and relationship mapping, query decomposition for complex questions. - **Smart Query Processing**: Automatic query decomposition, hybrid vector + graph-based retrieval, retrieval quality evaluation, context-aware response synthesis, conversation memory with graph context. - **Chat & Session Management**: Persistent chat sessions, context-aware follow-up questions, session renaming and management, message tracking with context preservation, multi-user support with isolation. - **Security & Authentication**: JWT-based authentication, secure password hashing with bcrypt, role-based access control, per-user data isolation, document access verification. - **Storage & Infrastructure**: S3-compatible object storage, PostgreSQL for structured data, Neo4j for graph relationships, concurrent file operations, efficient batch processing. ### Tech Stack: Python ### Why It Matters: KnowFlow is a comprehensive solution for intelligent document processing and querying, leveraging advanced semantic search and knowledge graph reasoning to enhance the efficiency and accuracy

### Overview: DocuMind is a local-first RAG platform that leverages FastAPI, ChromaDB, and Ollama for technical document libraries. It provides a RESTful API for querying and managing documents, with FastAPI handling the application layer and Ollama for inference. The system is designed for efficient retrieval and generation, with a focus on grounding first and explicit provenance. ### Key features: - **FastAPI**: The main backend for handling requests and responses. - **ChromaDB**: Stores and retrieves documents using vector embeddings. - **Ollama**: The primary model for inference, used for generating responses. - **Configuration**: Settings and logging configurations. - **Deployment**: Single library index per deployment, ensuring single-tenant baseline. ### Tech stack: - Python - FastAPI - ChromaDB - Ollama ### Why it matters: DocuMind simplifies the process of managing and querying technical documents, making it easier for developers to access and understand complex information efficiently. The use of FastAPI and Ollama ensures that the system is both efficient and reliable, while the configuration and deployment strategies ensure that it can be easily extended and scaled. ### Repository: - **Repository**: cdtalley/DocuMind - **Stars**: 7 - **Description**: Local-first RAG platform for technical document libraries. Features FastAPI, ChromaDB, and active retrieval (FLARE) powered by Ollama. ### Table of contents: 1. System overview 2. Design principles 3. Repository layout 4.

Overview: This project demonstrates how to integrate Pinecone vector database with OpenAI embeddings in Scala using the Pinecone Scala Client and OpenAI Scala Client. Key features: - Pinecone vector database integration - OpenAI embeddings integration - Scala-based implementation Tech stack: Scala Why it matters: This project serves as a practical example for developers looking to utilize Pinecone and OpenAI for vector database and text embedding functionalities in Scala applications.

### Overview: This is a seed project demonstrating how to use the Pinecone Scala client. It includes seven single-function demos and two complex examples for Pinecone Index and Vector, suitable for both pod-based deployments and general use. ### Key features: - List Indexes - Create Index - Describe Index - Delete Index - Upsert Vectors - Query Vectors - Query Vectors By Filter - Pinecone Index Long Demo - Pinecone Vector Long Demo ### Tech stack: Scala ### Why it matters: This project serves as a practical guide for developers looking to integrate Pinecone into their Scala applications, showcasing various functionalities and examples of how to interact with Pinecone's API.

Overview: This guide provides a field-tested, multi-GPU vLLM tensor-parallel setup on an Intel Arc Pro B70 (Battlemage) workstation, using bare-metal vLLM and oneAPI 2025.3. The setup includes fixes for driver setup, multi-root USM, and error reporting, making it accessible to both humans and AI coding assistants. Key features: - Field-tested vLLM setup on Intel Arc Pro B70 - Multi-GPU tensor-parallel (TP=2/TP=4) configuration - Bare-metal vLLM setup with Intel XPU runtime - Fixes for driver setup, multi-root USM, and error reporting - OneAPI 2025.3 driver and vLLM version 0.20.2 Tech stack: Shell Why it matters: This guide offers a practical solution for running modern LLMs on a new generation of hardware, providing fixes for common issues and making the setup accessible to both humans and AI coding assistants. Repository: chriswagner-ai/intel-arc-b70-vllm-multi-gpu GitHub stars: 7 GitHub description: Field-tested guide: multi-GPU vLLM tensor-parallel (TP=2/TP=4) on Intel Arc Pro B70 (Battlemage BMG-G31, Xe2) on Linux. Driver setup (xe force_probe=e223), bare-metal vLLM + oneAPI 2025.3, the

### Overview: SmartPortAgent 是一个基于大模型（LLM）和RAG技术的智能口岸物流助手，旨在解决外贸人员在货物出口过程中遇到的“数据碎片化”和“异常状态难解读”痛点。 ### Key features: - **🤖 智能任务规划 (Agent)**: 自动编排任务，查询集装箱、报关和船期状态，并结合内置的海关法规知识库诊断异常。 - **📚 专家知识库 (RAG)**: 内置宁波口岸查验流程与监管政策，利用FAISS + m3e-base实现本地化高精检索。 - **🔌 全链路模拟**: 包含集装箱进港、报关放行、船舶截关时间等完整的模拟业务数据。 - **💡 智能风控**: 综合船期与查验进度，自动计算风险等级并给出“预漏装”等专业建议。 ### Tech stack: - **开发语言**: Python 3.10+ - **依赖管理**: `uv` (极速包管理) - **LLM 框架**: LangChain v0.1 - **向量检索**: FAISS (CPU) + HuggingFace (m3e-base) - **前端界面**: Streamlit ### Why it matters: 通过结合大模型和RAG技术，SmartPortAgent 为外贸人员提供了一个高效、准确的工具，帮助他们快速处理复杂的物流问题，减少信息碎片化和决策困难，从而提高通关效率和

Overview: CyborgDB is a JavaScript/TypeScript client SDK for CyborgDB, an encrypted vector database that self-hosts in your environment. It performs vector search on encrypted data and is designed for self-hosted use, ensuring Cyborg Inc. never receives your keys, vectors, or queries. Key features: - Encryption-in-use: Search runs over ciphertext, only the final top-k candidates are decrypted in the service. - Encrypted ANN: Disk-backed encrypted DiskIVF index. - Filters on encrypted metadata: Combine vector similarity with equality and range predicates in a single request. - BYOK / HYOK: Wrap per-index keys with AWS KMS or AWS Secrets Manager, or hold the key client-side. - Per-tenant key isolation: Per-index, per-user keys with cryptographic RBAC. - TypeScript-first API: Complete type definitions and IntelliSense for JavaScript and TypeScript apps. Tech stack: TypeScript Why it matters: This SDK simplifies the integration of CyborgDB into JavaScript and TypeScript applications, providing a self-hosted encrypted vector database solution that ensures data privacy and security. Repository: cyborginc/cyborgdb-js GitHub stars: 7 GitHub description: JS & TS Client SDK for CyborgDB: The Encrypted Vector Database

Overview: This project demonstrates how to integrate Weaviate, a vector database, with TypeScript and OpenAI's embedder to create and manage vectors for datasets. It includes example datasets scraped from Y Combinator and a Lord of the Rings character dataset, showcasing how to populate and query the database. Key features: - Integration of Weaviate with TypeScript - Use of OpenAI's embedder for vector creation - Example datasets: Y Combinator companies, Lord of the Rings characters - Populating the Weaviate database with example datasets - Viewing and querying data via GraphQL API - Visualizing data using Tensorflow's Embedding Projector Tech stack: TypeScript, Weaviate, OpenAI's embedder Why it matters: This project provides a practical example for developers looking to integrate vector databases with TypeScript, showcasing the benefits of using vector databases for efficient similarity searches and indexing of high-dimensional data.

### Overview: OnLLM is a platform designed to run Large Language Model (LLM) or Simultaneous Learning Model (SLM) models using OnnxRuntime directly on low-end devices such as low-power computers and mobile phones. It leverages Kivy and open-source technologies to ensure a cross-platform experience. ### Key Features: - An inbuilt chatbot to interact with the AI model(s). - A local offline Read-Any-When (RAG) feature to query on selected documents (PDF/DOCX) directly in the chatbot. - No trackers, no ads, and no data collection, ensuring complete privacy. ### Tech Stack: - Python ### Why it Matters: OnLLM provides a convenient and accessible way to utilize Large Language Models, making AI technology more accessible to everyone, especially those on low-end devices.

### Overview: Marrow is an MCP server designed to provide persistent memory for AI coding agents, ensuring they can pick up where they left off across sessions. ### Key Features: - **Session State Persistence**: Keeps agent's workspace and session state intact. - **Task Backlog**: Supports semantic search and automatic dependency unblocking for task tracking. - **Versioned Documents**: Stores and rolls back documents for versioned project artifacts. - **Semantic Code Search**: Facilitates finding code by meaning rather than filename. ### Tech Stack: Python ### Why It Matters: Marrow ensures that AI coding agents retain their context and progress between sessions, making them more efficient and reliable in complex projects. ### Repository: desikai-lab/Marrow ### Primary Language: Python ### GitHub Stars: 7 ### GitHub Description: Marrow is an MCP server providing persistent task tracking, versioned docs, and semantic code search — so any agent, in any session, can pick up exactly where the last one left off. ### Quickstart (5 minutes) #### Option A: Docker (Recommended) - **Prerequisites**: Docker and Docker Compose. - **Instructions**: 1. Clone the repository: `git clone https://github.com/desikai-lab/Marrow.git` 2. Navigate to the project directory: `cd Marrow` 3. Run the Docker Compose file: `docker-compose up` #### Option B: Glama (Hosted No Shell Access) - **Instructions**: 1. Clone the repository: `git

### Overview: Serve **nvidia/Qwen3.8-Flash-Next-FP4-Local** from one NVIDIA DGX Spark through an OpenAI-compatible API. ### Key Features: - **nvidia/Qwen3.8-Flash-Next-FP4-Local**: A pre-trained model optimized for DGX Spark. - **DGX Spark**: A high-performance GPU cluster with unified memory. - **vLLM (OpenAI-compatible API)**: The API endpoint for interacting with the model. - **One-command Recipe**: A script that automates the setup and start of the server. ### Tech Stack: - Shell: The primary language used for scripting and executing commands. ### Why It Matters: This recipe simplifies the process of deploying a high-performance model optimized for DGX Spark, allowing users to quickly access the model without needing to manage the underlying infrastructure. The use of an OpenAI-compatible API ensures compatibility with existing tools and workflows, making it easier to integrate into existing projects and environments.

Overview: LLMRoboFund is a powerful chatbot equipped with Multi-document RAG technology, designed to provide real-time investment research and financial data through a seamless interaction with an LLM. This chatbot leverages Retrieval-Augmented Generation to update its knowledge base with up-to-date documents from various sources, including the Turkey Electronic Fund Trading Platform and Public Disclosure Platform. Key features: - Real-time investment research and financial data - Multi-document RAG technology - Retrieval-Augmented Generation method for knowledge base updating - Access to diverse financial information and documents Tech stack: Jupyter Notebook Why it matters: LLMRoboFund simplifies the investment research process by eliminating the need for dedicated research hours, making it an efficient tool for informed investment decisions. Repository: dfavenfre/LLMRoboFund GitHub stars: 7 GitHub description: LLMRoboFund is a powerful chatbot empowered with Multi-document RAG. The chatbot is equipped with RetrievalQA and SQL Agents to ease the investment research.

### Overview: ImageBench is a benchmark suite for text-to-image models, featuring 50+ models and 192 prompts. The frozen prompts and specialized judge questions ensure consistency and fairness across models. The VLM judges evaluate models based on concrete test-specific questions, providing a binary PASS/FAIL verdict for every image. The results are published on the website, ensuring reproducibility and transparency. ### Key features: - **Fixed Benchmark**: Same 192 generation prompts and specialized judge questions across all models. - **Capability Probe**: Targets text rendering, counting, negation, hands, physics, reflections, and world knowledge. - **Automated Grading**: VLM judges evaluate models based on a binary PASS/FAIL verdict for every image. - **Everything is Published**: Every generated image, every judge verdict, and every aggregated score are publicly available. ### Tech Stack: - Python - Jupyter Notebook - GitHub Actions ### Why it Matters: - Ensures consistency and fairness across models. - Provides a concrete, test-specific evaluation metric. - Facilitates reproducibility and transparency in the benchmarking process. ### Repository: Repository: dh7/image-bench-ai GitHub Stars: 7 GitHub Description: ImageBench — text-to-image benchmark suite. 50+ models, 192 prompts, VLM judges. See results at https://imagebench.ai ### Repo Layout: - `imagebench-v1/`: Contains the frozen challenge suite (192 rows, one per prompt variant). - `imagebench

Overview: Ekip is a tool designed to streamline the Claude Code process, ensuring that agents can efficiently take over the workload while Claude focuses on planning and reviewing. Key features: - Customizable project memory - Integrated LLM wiki - Agent-based workflow Tech stack: JavaScript Why it matters: Ekip simplifies the Claude Code process, making it more efficient and effective for teams, while maintaining the high quality and expertise of the Claude Code sessions.

Overview: KIAgent is a personal AI brain built on open-source technology, designed to run locally and speak Model Context Protocol (MCP). It ingests data from your personal sources and serves it to AI assistants, providing a contextually aware AI experience. Key features: - Local, open-source, and MCP-speaking - Ingests mail, documents, chats, and notes - Runs on your own machine, with on-device inference using LLaMA.cpp - A platform that integrates with third-party services and local filesystem Tech stack: TypeScript Why it matters: KIAgent empowers individuals to build their own AI brains, providing a contextually aware AI experience that can act with real-world context without leaving their machine. Repository: edjafarov/kiagent-core GitHub stars: 7

Overview: JIT LoRA is a real-time continuous model learning system that modifies a running language model's weights mid-conversation on consumer Apple Silicon hardware. It uses MLX-native autograd for gradient-based LoRA adaptation and updates weights after every response. Key features: - Real-time continuous model learning - Gradient-based LoRA adaptation - Updates weights after every response - Uses MLX-native autograd for efficient gradient-based LoRA adaptation - Updates weights via background backpropagation Tech stack: Python Why it matters: JIT LoRA allows for real-time continuous model learning during conversations, improving the model's adaptability and relevance in dynamic interactions. This system is particularly useful for applications requiring immediate feedback and continuous improvement, such as AI assistants or conversational agents. Repository: eelbaz/jit-lora GitHub stars: 7 GitHub description: Real-time continuous model learning during conversation. The model rewrites its own weights while you talk to it.

### Overview: The jw-agent-toolkit is a comprehensive toolkit for accessing and managing content from JW: JW.org, Wol.jw.org, and related platforms. It includes a monorepo of Python packages that cover HTTP access, offline parsers, hybrid RAG, procedural agents, fine-tuning with doctrinal alignment, interpretability, multimodal generation, and a local MCP server with 129 tools and 28 REST endpoints for client apps. ### Key features: - **HTTP Access**: Provides direct access to JW.org and Wol.jw.org content. - **Offline Parsers**: Supports offline parsing of JW content. - **RAG Hybrid**: Integrates both fine-tuned and unsupervised RAG models. - **Procedural Agents**: Enables procedural agents for dynamic content generation. - **Fine-Tuning**: Supports fine-tuning with doctrinal alignment. - **Interpretability**: Implements interpretability techniques like probing lineal, steering, and Gemma/Qwen Scope. - **Multimodal Generation**: Generates content with guardrails. - **Second-Brain**: Utilizes GraphRAG for enhanced multimodal understanding. - **MCP Server**: A local MCP server with 129 tools and 28 REST endpoints for client apps. ### Tech stack: Python ### Why it matters: The jw-agent-toolkit is a critical tool for developers and content creators looking to access and manage JW content efficiently. Its comprehensive features and robust infrastructure make it an essential resource for anyone working with JW's content ecosystem.