Projects

Overview: This repository, named **RewardLM**, is designed to reward a language model with pancakes 🥞, providing a comprehensive solution for optimizing generative models through reinforcement learning and fine-tuning techniques. The repository includes three main modules: Reinforcement Learning from AI Feedback (RLAIF), Model fine-tune, and a Toxicity Meter for measuring the toxicity of the responses. Key features: - 🥞 Reward a generative LM using the DIALCONAN dataset with PPO reinforcement learning. - 👨🏼‍🏫 Fine-tune any generative model on the DIALCONAN dataset. - ⚖️ Toxicity Meter for assessing the toxicity of generative model outputs. Tech stack: Jupyter Notebook Why it matters: This repository offers a practical way to enhance the performance of language models by rewarding them with positive feedback, which can significantly improve their quality and relevance in various applications, such as customer service, content creation, and more. GitHub stars: 11

Overview: This project is a Cosmos API starter application designed to analyze files and interact with large volumes of textual data through a chat interface. Key features: - Upload and manage file indexes - Generate embeddings using the Cosmos API - Interactive chat interface for querying indexed content - Sourced responses using RAG techniques - Modern, responsive UI with Tailwind CSS and shadcn/ui components - Real-time data fetching powered by TanStack Query Tech stack: TypeScript, Next.js 15, FastAPI, Cosmos Client by Delos Intelligence Why it matters: Ideal for researchers, content managers, and professionals working with large volumes of textual data who require swift and precise information retrieval.

Overview: BikoDB is a Rust-based high-performance multi-model knowledge graph database that supports SQL, Cypher, and Gremlin queries. It includes 11 graph algorithms, AI/ML with HNSW vectors, and a REST API. The project is known for its zero garbage collection and blazingly fast performance. Key features: - Multi-model graph runtime - SQL, Cypher, and Gremlin query support - 11 graph algorithms - AI/ML with HNSW vectors - REST API - Zero garbage collection - Blazingly fast performance Tech stack: Rust Why it matters: BikoDB is a powerful and efficient database solution for graph, document, and vector workloads, making it suitable for applications requiring complex querying and AI/ML capabilities. Repository: DioCrafts/BikoDB GitHub stars: 10 GitHub description: ⚡ High-performance multi-model knowledge graph database written in Rust. SQL, Cypher & Gremlin queries · 11 graph algorithms · AI/ML with HNSW vectors · REST API · Zero GC · Blazingly fast.

Overview: Sudoku4LLM is a Sudoku dataset generator designed for training and evaluating reasoning in Large Language Models (LLMs). It offers customizable puzzles, difficulty levels, and 11 serialization formats to support structured data reasoning and Chain of Thought (CoT) experiments. Key features: - Customizable puzzles - Multiple difficulty levels - 11 serialization formats Tech stack: Python Why it matters: Sudoku4LLM provides a systematic, scalable, and flexible Sudoku dataset generator tailored for training, fine-tuning, and evaluating reasoning LLMs. It supports experiments in structured data reasoning, Chain of Thought (CoT) prompting, and tabular data serialization, making it an essential tool for advancing reasoning-focused LLMs.

### Overview: BERT模型在IMDb评论数据集上进行二分类（情感极性分析）时，取得了88.93%的准确率，这是使用官方demo超参数设置的结果。然而，通过调整超参数，模型的准确率进一步提升到92.48%，这表明模型的性能可以通过适当的参数调整得到显著改善。 ### Key features: - 使用BERT模型对IMDb评论数据集进行二分类（情感极性分析） - 包含完整数据集下的超参数调整结果 - 提供不同epoch、学习率、最大序列长度和数据集大小的测试结果 - 包括机器学习方法（朴素贝叶斯和线性支持向量机）的测试结果 - 使用LSTM和CNN模型进行对比测试 ### Tech stack: Jupyter Notebook ### Why it matters: BERT模型在IMDb评论数据集上的应用展示了深度学习在自然语言处理领域的强大能力，特别是在情感分析任务中的准确性和鲁棒性。通过调整超参数，可以进一步优化模型性能，从而提高实际应用中的效果。

Overview: EigenDB is a fast in-memory vector database written in Go, utilizing the Hierarchical Navigable Small Worlds (HNSW) algorithm for efficient querying. It offers a REST API and Python SDK, making it suitable for various applications. Key features: - In-memory vector database - Supports 2D vector space - REST API for integration - Python SDK for seamless integration Tech stack: Go, Faissgo (Go library for HNSW algorithm), Python. Why it matters: EigenDB is designed for real-time applications requiring high-speed vector queries, making it ideal for use cases such as recommendation systems, image search, and more. Repository: Eigen-DB/eigen-db GitHub stars: 11 GitHub description: The blazingly fast in-memory vector database 🚀

### Overview: This project is a desktop GUI application built with Python and DelphiFMX, utilizing the Replicate API to generate images from text prompts. The user-friendly interface allows for intuitive input and real-time feedback, making it a valuable tool for generating images based on detailed text descriptions. ### Key Features: - **Text Input**: Users can input text prompts to generate images. - **Image Display**: The generated image is displayed within the application. - **Real-Time Status**: Provides real-time updates on the image generation process. - **Error Handling**: Handles errors and displays process feedback. - **DelphiFMX GUI**: Built for cross-platform compatibility. - **Replicate API Integration**: Interacts with the Flux 1.1 Pro model hosted on Replicate. ### Tech Stack: - Python for application logic and API integration. - DelphiFMX for the desktop GUI. - Replicate API for text prompt generation. ### Why It Matters: This project offers a practical solution for generating images based on text prompts, making it useful for various creative and design applications. The integration of Python and DelphiFMX ensures a versatile and user-friendly interface, making it accessible across different platforms.

### Overview: Forge kernels are a set of optimized Triton kernels designed for faster, leaner fine-tuning of large language models (LLMs). These kernels are seamlessly integrated into Hugging Face models via the `forge.patch(model)` function, providing significant performance improvements without requiring model rewrites or changes to training scripts. ### Key features: - **One-call patching:** `forge.patch(model)` swaps the kernels into a live model in place, ensuring no regressions. - **Per-call shape guard:** Each shape-sensitive kernel falls back to the original forward when a call is too small for fusion to pay off. - **Exact, not approximate:** Forward and backward match eager within fp32/fp64 tolerances, with 660 tests passing on an A100-80GB. - **Evidence-backed:** Every speedup below maps to a committed CSV/JSON result file, detailing where each kernel loses. - **FSDP2-safe:** Patched forwards keep working when weights become sharded DTensors, verified for equivalence on 2 GPUs. ### Tech stack: Python ### Why it matters: Forge kernels provide a significant performance boost for LLM fine-tuning, making it easier to train models on smaller hardware or with limited resources. The kernels are designed to be compatible with existing infrastructure and frameworks, ensuring minimal disruption to existing workflows.

### Overview: CARVE is an open-source Jev-like model family utilizing Qwen3.8-27B, designed for one-token decisions. It offers typed outputs through a TypeSafe-compatible API, ensuring honest semantics and raw restricted candidate-token preferences. ### Key features: - **One-Token Decisions**: No free-form response generation or JSON parsing. - **Typed Outputs**: Noul, Choice, and Score responses through a familiar System One contract. - **Reproducible Evidence**: Public protocols, frozen revisions, raw summaries, ablations, and negative results live beside the code. - **Honest Semantics**: Every response identifies its probability level instead of presenting raw softmax as calibrated confidence. ### Tech stack: - Python 3.11+ ### Why it matters: - **One-token decisions**: No free-form response generation or JSON parsing. - **Typed outputs**: Noul, Choice, and Score responses through a familiar System One contract. - **Reproducible evidence**: Public protocols, frozen revisions, raw summaries, ablations, and negative results live beside the code. - **Honest semantics**: Every response identifies its probability level instead of presenting raw softmax as calibrated confidence. ### Repository: - **GitHub Repository**: [FuturixAI-and-Quantum-Works/carve-jev](https://github.com/FuturixAI-and-Quantum-Works/carve-jev) - **Stars**: 11 - **Description**: Open-source Jev-like model family built on Qwen3.8-27B,

### Overview: Paparats MCP is a self-hosted MCP server that integrates Claude Code, Cursor, Codex, GitHub Copilot, and other AI coding agents, enabling semantic code search across multiple repositories. It provides precise context and reduces new token usage by up to 23% and 25% faster execution times. ### Key Features: - **Semantic Code Search**: Searches across repositories for specific functions or changes. - **Meaningful Queries**: Answers queries with precise context, reducing new token usage. - **Local Embedding Models**: Runs on local Docker containers, keeping code with you. - **GitHub Integration**: Supports various MCP clients and integrates with Claude Code, Cursor, Codex, and GitHub Copilot. - **Large Repository Support**: Works efficiently on large, legacy codebases spread across multiple repositories. ### Tech Stack: - **Docker**: Runs the MCP server on your machine or server. - **Node.js 24+**: Required for the CLI. - **LLama.cpp**: Local models with Metal for macOS, CPU-only for Docker on other platforms. ### Why It Matters: - **Efficiency**: Faster execution times and reduced new token usage. - **Contextual Answers**: Precise context and meaningful queries. - **Versatility**: Integrates with multiple AI coding agents and supports various MCP clients. ### Repository: IBazylchuk/paparats-mcp - **Stars**: 11 - **Primary Language**: TypeScript - **GitHub Description**: A self-hosted MCP server for semantic code search with Qdrant and

### Overview: SpellX is an end-to-end MLOps platform designed for spell correction, covering data versioning, reproducible model training, containerized serving, automated CI/CD, and configuration-driven architecture. ### Key Features: - **Reproducible ML Pipelines**: Four DVC-managed stages for data ingestion, validation, training, and evaluation. - **Containerized Model Serving**: A Flask API that loads a trained model and serves spell correction over HTTP. - **Automated CI/CD**: GitHub Actions workflows for building Docker images, pushing signed images to GitHub Container Registry, and validating DVC pipelines. - **Config-Driven Architecture**: YAML-based configuration for separating pipeline logic from parameters. ### Tech Stack: Python ### Why It Matters: SpellX demonstrates the practical implementation of MLOps principles in a real-world NLP task, covering data versioning, reproducible model training, containerized serving, automated CI/CD, and configuration-driven architecture.

### Overview: AnyLoom: AnythingLLM Local AI Agentic Stack **A fully local, multi-agent AI system that gives you ChatGPT-level intelligence with complete privacy and control over your data.** ### Key features: - 🔒 **100% private** — No data leaves your machine. No API keys. No subscriptions. - 🧠 **Advanced reasoning** — Qwen3-30B MoE (30.5B params, 3.3B active) with hybrid thinking mode - 📚 **Hybrid RAG search** — Finds YOUR information better than pure vector search (dense + sparse retrieval) - 🤖 **Multi-agent swarm** — DyTopo coordination routes complex tasks to specialized agents that collaborate, with optional RAG context pre-fetch for domain grounding - 🛡️ **Reliability hardening** — Checkpoint crash recovery, deny-first policy enforcement, deterministic output verification, stalemate detection - 🛠️ **8 MCP servers** — Memory knowledge graph, web search, file operations, sequential thinking, RAG search, multi-agent swarm, system diagnostics - 🐋 **Docker-first architecture** — One command to start/stop everything. Auto-restart. Zero networking hassles. - 💬 **AnythingLLM UI** — Clean interface for chat, document Q&A, and workspace management ### Tech stack: Python ### Why it matters: If you need ChatGPT-level capability for sensitive work, AnyLoom gives you near the same intelligence without the privacy

Overview: This repository contains curated copy-paste prompt templates for AI product photography, focusing on e-commerce listings, detail shots, lifestyle scenes, marketing posters, and product image-to-video ads. The templates follow a consistent structure and work with various image models. Key features: - White background / e-commerce listing - Product detail close-ups - Lifestyle scenes - Marketing posters - Food and drink - Product image to ad video Tech stack: The repository uses GPT Image, Nano Banana, and Midjourney. Why it matters: These templates provide a consistent and effective way to generate high-quality AI product photography, ensuring that the product details and environment are accurately represented.

Overview: NeuralForge is a macOS application that allows fine-tuning of transformer models directly on Apple Silicon using the Apple Neural Engine (ANE). It supports on-device training, text generation, and cloud backup, making it ideal for developers and data scientists working with AI models on Apple devices. Key features: - On-device training of large language models (LLMs) using the Neural Engine - Native macOS app with a SwiftUI dashboard for live loss charts, project management, and integration with the menu bar - Support for Low-Rank Architecture (LoRA) with configurable rank and target layers - Text generation with interactive inference and temperature sampling - Multi-shard loading, train/val split, and tokenization - LR scheduler with cosine annealing with warmup - Export options for GGUF, CoreML, and llama2c formats - Distributed training via Bonjour for multi-Mac cluster support - Cloud backup with S3 and iCloud checkpointing - Enterprise audit logging and compliance reports - INT8 and INT4 weight quantization with calibration - Full configuration presets save/load support - Drag & drop data integration - Webhook notifications for Slack/Discord alerts on training events - MLX backend alternative Metal GPU backend for broader Mac compatibility Tech stack: Swift Why it matters: NeuralForge simplifies the process of fine-tuning large AI models directly on Apple Silicon, leveraging the powerful Neural Engine for efficient training and inference, while ensuring data stays on the device. This tool is particularly valuable for developers and data scientists who need to work with AI models on

### Overview: Tune-A-VideKO is a Korean-stable-diffusion-based one-shot video tuning project that utilizes the Tune-A-Video framework. It connects the Korean version of stable-diffusion with Tune-A-VideKO to create visually engaging and realistic videos. ### Key Features: - **Tune-A-VideKO-v1-5**: A Korean-stable-diffusion-based video tuning model. - **Tune-A-VideKO-anything**: A versatile model that can be applied to any input video. - **Tune-A-VideKO-disney**: A specialized model for creating Disney-style videos. ### Tech Stack: - **One-shot video tuning**: Utilizes the Tune-A-Video framework. - **Stable Diffusion**: The Korean version of the stable-diffusion model. - **Diffusers**: The Hugging Face library for model training and inference. ### Why it Matters: Tune-A-VideKO provides a unique and efficient way to create visually appealing videos by leveraging the strengths of stable-diffusion and Tune-A-Video, resulting in high-quality and engaging video outputs.

Overview: This project explores methods to enhance Large Language Models (LLMs) by integrating custom data and the ability to call functions, aiming to improve their performance for specific domain-related tasks. Key features: - Function Calling - Vector Databases & Retrieval Augmented Generation (RAG) Tech stack: Jupyter Notebook Why it matters: This project aims to provide hands-on experience with technologies and concepts related to LLM augmentation, ultimately improving the performance of LLMs for specific domain-related tasks. Repository: MichiganDataScienceTeam/llm-augmentation GitHub stars: 11 GitHub description: Applying LLMs, augmented with information retrieval (RAG) and function calling, to a variety of tasks in different domains.

### Overview: Serverless picoLLM is a project that enables the inference of large language models (LLMs) using AWS Lambda, making it accessible and cost-effective for developers. ### Key Features: - **LLM Inference on AWS Lambda**: Runs large language models (LLMs) in a serverless environment, leveraging AWS Lambda for efficient execution. - **Picovoice Console Integration**: Utilizes the Picovoice Console for managing and deploying models. - **Model Compatibility**: Supports various models, including Phi2, Llama3, and others, as long as they are chat-enabled and fit within AWS Lambda's code size and memory limits. ### Tech Stack: Python ### Why it Matters: Serverless picoLLM simplifies the deployment and execution of large language models, making them accessible and cost-effective for developers. This project is particularly useful for integrating LLMs into serverless architectures, reducing the need for on-premises servers and lowering operational costs.

Overview: This repository contains TensorFlow Serving demos, showcasing various use cases and best practices for deploying models in production. Key features: - Deploying models to production - Real-time predictions - Model serving with TensorFlow Serving - Integration with Jupyter Notebooks Tech stack: Jupyter Notebook Why it matters: These demos provide practical examples and insights into deploying TensorFlow models efficiently, making them valuable resources for developers looking to integrate TensorFlow Serving into their projects. Repository: Rishit-dagli/TFServing-Demos GitHub stars: 11 GitHub description: TF Serving demos README excerpt:

Overview: Docker containers on macOS can now leverage the full Metal GPU capabilities, thanks to the NVIDIA Container Toolkit and docker_mlx_cpp. Key features: - Zero CUDA - Zero NVIDIA - Metal GPU access - 100+ GPU operations - LLM inference - Training - Image generation - Audio processing - Embeddings Tech stack: Python Why it matters: This tool allows developers to run Metal GPU-accelerated applications on macOS without the need for CUDA, making it ideal for AI and machine learning workloads on the Apple Silicon ecosystem. GitHub stars: 11

Overview: This repository provides hands-on Azure Kubernetes (AKS) demos for KAITO, RAGEngine, GPU inference, and Azure Key Vault, designed to be used in technical presentations and learning environments. Key features: - Provision AKS and related Azure resources - Run KAITO model workspaces - Deploy and test RAGEngine - Expose services via ingress - Build practical chat/RAG demo workflows Tech stack: Python Why it matters: This repository offers a practical way to learn and demonstrate Azure Kubernetes, KAITO, RAGEngine, GPU inference, and Azure Key Vault, making it valuable for technical presentations and educational purposes. Repository: RoyKimYYZ/aks-demos GitHub stars: 11 GitHub description: Hands-on AKS demos for KAITO, RAGEngine, GPU inference, and Azure Key Vault.