Projects

### Overview: Autosynth is an agentic synthetic-data generation framework inspired by Meta FAIR's Autodata/Agentic Self-Instruct paper. It generates structured examples using an LLM loop that proposes, checks, solves, and scores its own examples. The framework is designed to be task-specific, allowing it to generate math problems, support-ticket data, document-grounded QA, and more. ### Key features: - **Task-Specific Behavior**: The same runtime can generate examples for various tasks, including math problems, support-ticket data, and document-grounded QA. - **Quality Checks**: Candidates are evaluated for obvious problems, and weak and strong solvers attempt the input multiple times. - **Feedback Loop**: Failed attempts become feedback for the next round, improving the quality of the generated examples. - **Acceptance Modes**: Acceptance can be set to rubric, verifiable, or judge, each with its own scoring criteria. ### Tech stack: Python ### Why it matters: Autosynth is an innovative tool for generating synthetic data, inspired by Meta FAIR's cutting-edge approach. It addresses the need for high-quality, structured examples that can be used for training and testing machine learning models. The framework's ability to generate diverse and relevant examples makes it a valuable asset in the development and validation of AI systems.

Overview: AuraLite AI is a tool designed to train and fine-tune large language models on a personal computer without the need for a GPU. It supports various models like Qwen3.5, Qwen2.5/3, Llama, Mistral, and SmolLM2, and allows users to fine-tune these models using QLoRA and train small models from scratch. Key features: - Fine-tuning Qwen3.5 (2026) and Qwen2.5/3 with QLoRA on a CPU - Training small models from scratch - Exporting models to formats compatible with Ollama, LM Studio, and transformers - Running models locally using PyTorch - Fine-tuning and training from scratch with a wizard interface - Providing a Windows app and command line interface - Supporting multiple languages (English and Russian) Tech stack: AuraLite AI uses PyTorch for training and runs on a computer without a GPU. Why it matters: AuraLite AI simplifies the process of training and fine-tuning large language models, making it accessible to users without the resources needed for GPU-based training. This tool is particularly useful for developers and enthusiasts who want to experiment with cutting-edge language models without the need for specialized hardware.

Overview: Fine-tuning Google's BERT model on text classification for sentiment analysis. Key features: - Fine-tuning BERT model on text classification tasks - Utilizing PyTorch and Keras for implementation - Transfer learning approach for sentiment analysis Tech stack: PyTorch, Keras Why it matters: This project demonstrates how to apply BERT for sentiment analysis, showcasing the effectiveness of transfer learning in text classification tasks.

Overview: This utility uses face recognition to automatically crop faces out of multiple images, making it easier to create datasets with faces cropped properly. Key features: - Automatically crops faces out of multiple images - Works even when there are multiple people in the pictures - Uses the face-recognition package for recognition Tech stack: Python Why it matters: This utility streamlines the process of cropping faces out of images, which can be time-consuming and error-prone when done manually.

Overview: This repository features a collection of generative AI applications designed to showcase the capabilities of AI technology across various domains. Key features: - Text summarization projects - Text generation projects - Text classification projects - Image synthesis models - Code generation tools Tech stack: Python, TensorFlow, PyTorch, Groq API, Google Gemini API, Hugging Face Transformers, Streamlit, Gradio, FastAPI or Flask Why it matters: This repository provides a platform for experimenting with and demonstrating the potential of generative AI across multiple domains, including text generation, image synthesis, and code creation. Repository: AryanKarumuri/Gen-AI-Projects Primary language: Jupyter Notebook GitHub stars: 6 GitHub description: This repository features a collection of generative AI applications designed to showcase the capabilities of AI across various domains. Explore projects that generate text, images, music, and code, each demonstrating innovative uses of generative technology. Contribute your own projects and experiment with the potential of AI! ## Table of Contents📑 - [Project Overview](#project-overview) - [Technologies Used](#technologies-used) - [API Key Generation Guide](#api-key-generation-guide) - [Deploying on Streamlit](#deploying-on-streamlit) - [Getting Started](#getting-started) - [Contributing](#contributing) - [Note](#note-on-model-output) - [License](#license) ## Project Overview📝 This repository showcases a variety of generative AI applications, each designed to illustrate

Overview: Floor planning using AI-driven models is showcased in this project, leveraging advanced fine-tuned models and pre-trained systems to generate precise and accurate floor plans. Key features: - AI-Generated Floor Plans - Model Fine-Tuning - API Integrations with Janus, Optimized_Cloudqi_Text_to_Image_Model, and Vertex AI - Custom Safety Mechanisms - Performance Optimization Tech stack: Jupyter Notebook Why it matters: This project demonstrates the application of AI in architectural design, showcasing how AI can enhance the accuracy and efficiency of floor plan generation, making it a valuable tool for architects and designers. Repository: Auth0r-C0dez/Neurathon-25 Primary language: Jupyter Notebook GitHub stars: 6 GitHub description: This is the project for the final round of neurathon 2025

Overview: SOLLOL is a performance-aware load balancer and orchestration layer designed for distributed Ollama deployments. It routes requests to the best available node based on GPU/CPU capacity and current load, ensuring optimal performance and reliability. Key features: - Intelligent routing that scores nodes based on success rate, latency, GPU availability, current load, priority, and task type. - Auto-discovery of Ollama nodes on the network. - GPU-aware routing to avoid routing heavy models to nodes without the necessary GPU. - Distributed execution with Ray for parallel request execution and Dask for batch processing. - Live dashboard for monitoring node health, latency, active applications, and GPU memory usage. Tech stack: Python, Ray, Dask, Redis, and Ollama. Why it matters: SOLLOL simplifies the management of distributed Ollama clusters, ensuring optimal performance and reliability through intelligent routing and observability. Repository: B-A-M-N/SOLLOL GitHub stars: 6 GitHub description: Super Ollama Load Balancer & Orchestration Layer

Overview: This project involves developing a multilingual LLM prompt-attack classifier for the RMIT SGS GenAI & Cybersecurity Hackathon 2026 competition. The classifier is fine-tuned on mmBERT-base and QLoRA-tuned Granite-2B, achieving a macro F1 score of 0.98241. The classifier classifies prompts as benign, jailbreak, or prompt injection, with a focus on fine-tuning models from the approved list and using 5-fold cross-validation for validation. Key features: The classifier is fine-tuned on mmBERT-base and QLoRA-tuned Granite-2B, achieving a macro F1 score of 0.98241. It classifies prompts as benign, jailbreak, or prompt injection, with a focus on fine-tuning models from the approved list and using 5-fold cross-validation for validation. Tech stack: The project uses Jupyter Notebook and fine-tuned models from the approved list, including mmBERT-base, QLoRA-tuned Granite-2B, and 60+ fine-tuning runs on 9 backbones. Why it matters: The project showcases the effectiveness of fine-tuning models and the importance of cross-validation in improving model performance, as demonstrated by the high private macro F1 score of 0.98241, which is 0.0036 better than the public score of 0.98111.

Overview: This repository contains 70 hands-on projects designed to help developers enhance their skills in Generative AI and related technologies. Ideal for beginners and experienced professionals alike, the projects cover a wide range of topics and provide practical applications. Key features: - Hands-On Projects: Over 70 projects covering various AI topics. - User-Friendly: No coding skills required. The platform guides you through each project. - Resource Materials: Includes detailed guides and Jupyter notebooks for reference. - Community Support: Join community forums for help and collaboration. Tech stack: Python, Jupyter Notebook, and other relevant tools. Why it matters: This resource is essential for anyone looking to improve their skills in AI and related technologies, providing practical experiences that can be directly applied in real-world projects. GitHub stars: 6 Repository: Brucewaynehatake/GenAI-Architect-70-Hands-On-Projects

### Overview: This project is a Khmer-language math AI solution for Cambodia's BacII exam, offering 108,332 worked solutions across 8 chapters. The model is fine-tuned on a Gemma 3 12B model and verified by SymPy, providing accurate and detailed answers in Khmer. ### Key features: - 108,332 worked solutions in Khmer, covering 8 chapters of the BacII exam. - Every answer is computed by SymPy, ensuring no language model wrote a solution. - A LoRA adapter trained on Gemma 3 12B model, fine-tuned for Khmer BacII style. - A local chat page that automatically checks answers against the dataset or SymPy. ### Tech stack: Python ### Why it matters: This project offers a comprehensive and accurate math solution for students preparing for the Cambodian BacII exam, ensuring they have access to detailed and verified answers.

### Overview: Reson is a fine-tuned version of LLaMA-2 7B HF, trained on a custom dataset of ~11,000 examples. It is designed as a decision-support and cognitive simulation system, focusing on extreme meta-cognition, multi-domain reasoning, and strategic adaptability. ### Key Features: - Multi-domain reasoning - Extreme meta-cognition - Strategic adaptability - Unconventional logic blending - Exploratory cognition ### Tech Stack: Python ### Why it Matters: Reson is a powerful tool for exploring complex cognitive processes and adaptive reasoning in a wide range of domains, pushing the boundaries of what traditional chatbots can achieve.

### Overview: DenseCore is a CPU inference server designed for large language models (LLMs) using the GGUF format. It provides a native C++ runtime and OpenAI-compatible API, making it suitable for high-performance inference on plain CPUs. ### Key Features: - **Native C++ Runtime**: Executes the GGUF model in native C++. - **OpenAI-Compatible API**: Supports chat and prompt completions, embeddings, and more. - **Production Readiness**: Includes readiness/liveness probes, Prometheus metrics, and bounded admission. - **Docker Image**: Provides a production-ready Docker container. - **Python SDK and Helm Chart**: Offers a Python SDK and a Helm chart for deployment. - **Typed Decisions**: Supports Laya GGUF with a native System One API (preview). - **Hardware Support**: Available for Linux x86-64, Linux Arm64, and experimental Apple and accelerator paths. ### Tech Stack: C++ and Go. ### Why It Matters: DenseCore significantly improves the performance of LLM inference on CPUs, achieving up to 39% faster first token latency compared to llama.cpp. This makes it ideal for applications requiring high throughput and low latency. ### Repository: Repository: [DenseAI/DenseCore](https://github.com/DenseAI/DenseCore) GitHub Stars: 6 GitHub Description: CPU inference server for GGUF MoE models.

Overview: LangBatch is a unified API for accessing Batch APIs from providers like OpenAI, Anthropic, Azure OpenAI, Vertex AI, and Bedrock, offering a streamlined approach to handling batch jobs and requests. Key features: - OpenAI format for requests and responses - Utilities for managing batch jobs: creation, starting, monitoring, retrying, and processing - Pipeline to convert incoming requests into batch jobs Tech stack: Python Why it matters: LangBatch simplifies the process of interacting with multiple Batch APIs, providing a single interface for developers to work with, thereby reducing complexity and improving efficiency. Repository: EasyLLM/langbatch Primary language: Python GitHub stars: 6 GitHub description: LangBatch is a unified API for accessing Batch APIs from providers like OpenAI, Anthropic, Azure OpenAI, Vertex AI, and Bedrock, offering a streamlined approach to handling batch jobs and requests.

Overview: Fine-tuning Microsoft Phi-4 Mini 3.8B on SEC 10-K financial Q&A using QLoRA — +69% ROUGE-L over base model. Live demo on Hugging Face. Key features: - Fine-tuning of Microsoft Phi-4 Mini 3.8B - Training on free Kaggle GPUs - 69% improvement in ROUGE-L over base model - Live demo on Hugging Face Spaces Tech stack: Jupyter Notebook Why it matters: This project fine-tunes a powerful language model on financial data to provide insights that are more accurate and relevant for financial analysts, making it a valuable tool for the finance industry. Repository: Emart29/phi4-finance-finetuning

Overview: MBASE NLQuery is a high-performance Text-to-SQL engine for PostgreSQL, designed for on-premise use. It leverages the Qwen2.5-7B-Instruct-NLQuery model from the MBASE SDK to convert natural language queries into SQL queries, executing them through the PostgreSQL client SDK. The engine is secure by default and can be configured to run offline without requiring external table information. Key features: - Offline Text-To-SQL generation/executor engine - Secure execution through 'generate_only' option - Easy to use single HTTP REST API endpoint - Supports HTTPS through OpenSSL - CUDA acceleration for GPU usage - Easy prompt cooking for text-to-text models - Simple web UI for testing purposes Tech stack: C++ Why it matters: MBASE NLQuery simplifies the process of converting natural language queries into SQL queries, making it easier for developers to integrate text-to-sql functionality into their PostgreSQL databases without the need for external table information. Its secure execution and easy-to-use REST API make it a valuable tool for developers working with PostgreSQL databases. GitHub stars: 6 Repository: Emreerdog/mbase_nlquery

Overview: CRAFTS is a multi-agent system designed for chemical process simulation, replacing a closed black-box simulation loop with role-adapted agents over an inspectable IDAES/Pyomo execution substrate. It includes seven agents and covers input understanding, intent routing, visual description, topology generation, specification generation, debugging, and optimization. Key features: - Multi-agent system - Role-adapted agents over an inspectable IDAES/Pyomo execution substrate - Seven collaborating roles - Three schema-critical roles fine-tuned for VisualGraphIR, TopologyIR, and SpecIR generation - Deterministic gates validate promoted artifacts before model construction, solving, and optimization - Supported IDAES flowsheets, reduced process models, grid and design optimization, and external dynamic simulation Tech stack: LangGraph, IDAES, Pyomo Why it matters: CRAFTS provides a collaborative and adaptive approach to chemical process simulation, enhancing the accuracy and reliability of process models by integrating human expertise and machine learning.

Overview: Vulnerability Bot is a local client/server system utilizing a vulnerability database to answer user queries about vulnerabilities. Key features: - Database integration for vulnerability information - Streamlit-based user interface for easy interaction - Local server for processing and updating queries - Python-based backend with support for various LangChain packages Tech stack: Python Why it matters: This project provides a practical tool for developers to quickly find and understand potential security vulnerabilities, streamlining the process of vulnerability management and enhancing security awareness.

Overview: This project optimizes the dequantization process for 4-bit NormalFloat (NF4) quantization using Triton GPU kernels, achieving up to 1.41x speedup over `bitsandbytes` C++ across various tensor sizes. Key features: - Single fused Triton kernel for efficient dequantization - Inline PTX assembly for optimized bit unpacking - L2 cache eviction for reduced memory access - BF16/FP16 output format - Compatible with PyTorch and Triton - >1.15x speedup over `bitsandbytes` C++ across all tensor sizes Tech stack: Python Why it matters: This project provides a significant performance improvement by eliminating CPU dispatch overhead and achieving consistent speedups, making it particularly useful for optimizing the dequantization process in applications that use 4-bit quantization. Repository: Griffith-7/nf4-triton-kernel GitHub stars: 6 GitHub description: An optimized, high-performance NF4 (NormalFloat 4-bit) dequantization Triton GPU kernel achieving up to 1.41x speedup over bitsandbytes.

### Overview: Saudi-Dialect-ALLaM is a code-only release designed for researchers to fine-tune ALLaM-7B-Instruct-preview for Saudi dialect generation. The project includes training, evaluation, inference scripts, and analysis notebooks. ### Key Features: - **Training Scripts and Configs:** LoRA fine-tuning scripts and a config to replicate the setup used in the research. - **Evaluation Suite:** Internal evaluation for ALLaM variants and external evaluation with a dialect classifier, chrF++, BERTScore, and diversity metrics. - **Inference Helpers:** Tools for generating prompts from a trained checkpoint. - **Notebooks:** End-to-end training, evaluation, and inference demos, as well as figure generation. - **Results:** JSON/CSV reports and figures used in the paper. ### Tech Stack: Python, Jupyter Notebook ### Why it Matters: The project showcases how to fine-tune language models for specific dialects, achieving high alignment scores and minimal leakage of non-target dialects. This can be crucial for applications requiring precise language generation in specific regions. ### GitHub Stars: 6 ### Repository: HasanBGit/Saudi-Dialect-ALLaM ### Primary Language: Jupyter Notebook ### GitHub Description: Official code for "Saudi-Dialect-ALLaM: LoRA Fine-Tuning for Dialectal Arabic Generation" — LoRA fine-tuning of ALLaM-7B-Instruct-preview for Saudi dialect (Hijazi & Najdi) with dialect-token control, achieving 84% Saudi alignment and

### Overview: JimboMesh Holler Server is an open-source AI inference server designed to replace cloud-based embedding calls with a local Ollama instance. It includes a Model Marketplace, Document RAG, and OpenAI-compatible APIs, all running on-premises for better data security and performance. ### Key Features: - Local Ollama server for embedding and inference. - Model Marketplace for easy model management. - Document RAG integration for enhanced search capabilities. - OpenAI-compatible API access. - Admin UI for centralized management. - Optional Mesh connectivity for seamless data transfer. ### Tech Stack: JavaScript, Node.js, Express, Docker, OpenAI, Qdrant, Ollama. ### Why It Matters: JimboMesh Holler Server ensures data remains on-premises, improving security and reducing costs associated with cloud-based services. The local Ollama server eliminates the need for API calls, making it faster and more efficient. The Model Marketplace and Document RAG features enhance the overall functionality of the system, making it a valuable tool for developers and data scientists alike. ### Repository: The repository is hosted on GitHub at [IngressTechnology/jimbomesh-holler-server](https://github.com/IngressTechnology/jimbomesh-holler-server). ### GitHub Stars: The repository has received 6 GitHub stars, indicating its popularity and community interest. ### Primary Language: The primary language used in the project is JavaScript. ### GitHub Description: This repository contains the JimboMesh Holler Server, an open-source AI inference server with Model