Projects

Overview: Open Server is an OpenAI API-compatible server designed to generate text, images, embeddings, and store them in vector databases. It includes a chat functionality. Key features: - Chat & Completions - Image Generation Tech stack: Python Why it matters: Open Server offers a server-based solution for generating text, images, and embeddings, making it accessible and efficient for various applications. Repository: Mj23978/OpenServer GitHub stars: 17

### Overview: GenQuest-RAG is an innovative system designed to generate insightful questions by leveraging the power of machine learning and Wikipedia as a knowledge repository. It integrates advanced natural language processing techniques to retrieve and augment context, producing more accurate, contextually relevant, and informative questions across various topics. ### Key Features: - Contextual Information Retrieval: Extracts relevant information from Wikipedia articles to generate more accurate and contextually rich questions. - Dynamic Question Generation: Produces high-quality questions tailored to specific topics or areas of interest. - Versatile Applications: Suitable for educational settings, content creation, and research, enhancing learning and comprehension. ### Tech Stack: Python, PyTorch, Transformers, Datasets, Evaluate, LangChain, Weaviate ### Why It Matters: By integrating RAG with Wikipedia, the project offers a powerful tool for generating questions that stimulate critical thinking and deepen understanding. The system's ability to retrieve and utilize relevant context from Wikipedia articles ensures that generated questions are more informed and contextually relevant, making it an invaluable resource for anyone looking to enhance their learning or research efforts. ### Repository: MohammedAly22/GenQuest-RAG ### GitHub Stars: 17 ### GitHub Description: A Question Generation Application leveraging RAG and Weaviate vector store to be able to retrieve relative contexts and generate a more useful answer-aware questions.

### Overview: Cynde is a Python framework designed to streamline the integration of large language models (LLMs) with the modern data science stack, such as Polars DataFrames and decision tree ensembles, for efficient and scalable data processing. The framework provides a set of modular and composable tools for tasks such as embedding generation, structured text generation, and predictive modeling, with a focus on leveraging the power of LLMs and serverless computing. ### Key Features: - **OpenAI API-Compatible Server Integration**: Cynde is compatible with any OpenAI API-compatible server, including popular LLMs like OpenAI's GPT models and self-hosted solutions like llama.cpp or OpenLLaMA. This allows users to leverage the power of LLMs while maintaining flexibility in their deployment options. - **Serverless LLM Integration**: Cynde seamlessly integrates with serverless computing platforms like Modal for massively scalable LLM integration. - **GPU Workload Optimization**: Cynde optimizes GPU workloads by leveraging continuous batching and optimized transformers code for inference using Flash Attention and Paged Attention on the most popular architectures. - **Observability and Logfire Integration**: Cynde integrates with Logfire, an observability platform, to provide powerful monitoring and insights into the performance and behavior of LLM-powered data processing pipelines. This enables users to gain deep visibility into the execution and performance of their workflows, making it easier to identify bottlenecks, optimize resource utilization, and ensure the reliability of their applications. ### Tech Stack: Python ### Why it Matters: Cynde offers a flexible and adaptable solution for

Overview: JAMIE, Joint variational Autoencoders for Multi-modal Imputation and Embedding, is an open-source tool designed to address the challenges of generating multi-modal data for single-cell multi-modal datasets, particularly focusing on missing modalities. This project aims to provide a machine learning model that can impute cell data and learn joint cross-modal latent embeddings, improving the understanding of cellular and molecular mechanisms. Key features: - Joint variational Autoencoders for multi-modal imputation and embedding - Open-source implementation with Python - Supports partially matched multi-modal data across modalities - Aggregates embeddings from matched samples to identify joint cross-modal latent embeddings - Utilizes opposite decoders for cross-modal imputation Tech stack: Python Why it matters: JAMIE addresses the limitations of traditional machine learning approaches that rely on fully matched multi-modal data, offering a solution for imputing missing modalities and learning specific latent embeddings for each modality, thereby improving the accuracy and understanding of cellular and molecular mechanisms in single-cell multi-modal datasets. Repository: Oafish1/JAMIE Primary language: Python GitHub stars: 17 GitHub description: Official implementation of JAMIE, Joint variational Autoencoders for Multimodal Imputation and Embedding

### Overview: ToolsFilter is a precision-driven tool recommendation system designed to filter relevant tools for conversational LLM responses, thereby reducing cost and improving model accuracy. ### Key Features: - Fetches only the necessary tools for the current conversation. - Saves cost by reducing the number of available tools. - Increases the precision of LLM responses. ### Tech Stack: - Python ### Why It Matters: - Modern LLMs face performance degradation when exposed to a large number of tools. - Research evidence shows an average 9.5% accuracy drop and 3.25× to 236.5× increase in input token volume. - The model struggles with instruction compliance and context mismatch when too many tools are available. - 🚀 Multi-Stage Search Pipeline: Semantic + BM25 + Cross-Encoder + LTR ranking - 🎯 High-Performance Results: Perfect P@1 and MRR across all search strategies - 🧠 Learning-to-Rank: XGBoost model with 46+ engineered features (NDCG@10: 0.975) - 🔧 OpenAI Function Calling Compatible: Flat tool structure following OpenAI specification ### Architecture: - 🚀 Search Pipeline Architecture: Multi-Stage Search Pipeline - 🎯 Search Strategies: Multi-Stage Search Pipeline ### Performance: - 📊 Latest Evaluation Results: Built-in framework with F1, MRR, NDCG@k metrics - 🔄 Optimization Roadmap: Continuous Improvement and Optimization ### Quick Start: -

Overview: This project is a demonstration of deploying a model using LLamacpp, a modal LLM, and deploying it as part of a Model as a Service (MaaS) initiative. The primary language used is Python. Key features: - Deploying a model using LLamacpp - Using Modal for easy deployment - Simplified inference endpoint creation Tech stack: Python Why it matters: This project showcases the simplicity of deploying a model using LLamacpp and deploying it on Modal, making it an ideal example for promoting the concept of Model as a Service (MaaS). Repository: OutofAi/ChitChat Primary language: Python GitHub stars: 17 GitHub description: This is the first part of a collection of templates for promoting the concept of Model as a Service (MaaS). It demonstrates the deployment of Mistral 7B Instruct v0.1 - GGUF using LLamacpp on Modal. README excerpt: This project provides a step-by-step guide on how to deploy a model using LLamacpp and deploy it as part of a Model as a Service (MaaS) initiative. It includes instructions on prerequisites, deployment, and inference endpoints.

Overview: A Rust library for embeddable vector search, supporting metadata filtering, persistence, and optional Python and WebAssembly interfaces. Key features: - Vector indexing and persistence - Metadata filters - Hybrid retrieval Tech stack: Rust Why it matters: This library provides a flexible and efficient solution for vector search, suitable for applications that need to store and search for similar records alongside metadata. Repository: PhilipJohnBasile/vecstore GitHub stars: 17 GitHub description: Embeddable vector search in Rust, with optional Python and WebAssembly interfaces. Alpha.

### Overview: Mirror Gradient (MG) is a gradient strategy designed to make multimodal recommendation models approach flat local minima, reducing inherent noise risk and information adjustment risk. This paper, accepted at the 2024 ACM Web Conference, proposes MG as a solution to address these issues in recommender systems. ### Key Features: - **Gradient Inversion**: Inverses the gradient signs during training to make models approach flat local minima. - **Modular Design**: Built on the MMRec framework, allowing easy integration into various recommendation models. - **Versatile Models**: Supports models like BM3, DRAGON, DualGNN, FREEDOM, GRCN, ItemKNNCBF, LATTICE, LayerGCN, MGCN, MVGAE, SELFCFED_LGN, SLMRec, VBPR, and others. - **Robust Training**: Simplifies training by reducing the risk of local minima, making it easier to achieve better performance. ### Tech Stack: - **Python**: The primary programming language used for development. - **MMRec**: A modern multimodal recommendation toolbox that serves as the foundation for the implementation. ### Why It Matters: - **Reduced Risk of Local Minima**: MG helps multimodal recommendation models avoid getting stuck in local minima, leading to more robust and accurate predictions. - **Improved Performance**: By making models approach flat local minima, MG can lead to better performance and reduced risk of overfitting. ### Repository: - **Repository**: Qrange-group/Mirror-Gradient -

Overview: This project is an AI-driven home energy management system that utilizes large language models (LLMs) to schedule and orchestrate various appliances, optimize electricity prices, and provide real-time reasoning and execution logs. It aims to explore agentic AI architectures in residential demand response scenarios. Key features: - ReAct Orchestration: Transparent decision-making using the Reasoning-Action pattern - Day-ahead pricing: Fetches electricity prices from the ENTSO-E for 15-minute resolution optimization - Multi-appliance coordination: Scheduling for appliances like a dishwasher, washing machine, and electric vehicle charger - Provider-agnostic: Works with multiple LLM providers (Cerebras, OpenAI, OpenRouter, local LLMs) - Real-time streaming: Observes agent reasoning through Server-Sent Events - Calendar integration: EV charging deadlines from Google Calendar - Optional device control: Experimental API integration for actual appliance control Tech stack: Python Why it matters: This project provides a foundational framework for exploring agentic AI architectures in residential demand response scenarios, enabling research in interpretable decision-making, flexible constraint handling, and systematic evaluation of large language models for home energy management. GitHub stars: 17 Repository: RedaElMakroum/agentic-ai-hems

### Overview: The Facial Recognition Based Attendance System is an automated solution designed to streamline attendance management in educational institutions. By utilizing advanced facial recognition technology, the system eliminates manual attendance processes, reduces proxy attendance, and provides real-time attendance tracking with comprehensive reporting capabilities and data warehousing. ### Key features: - **Real-time Facial Recognition using FaceNet deep learning model with MTCNN face detection** - **Multi-user Admin Panel with role-based access control for system management** - **Department-wise Student Management with hierarchical organization** - **Automated Email Notifications for registration and attendance reports** - **Google Sheets Integration for live attendance logging and data export** - **Vector Database (Qdrant) for fast and accurate face matching using cosine similarity** - **Databricks Data Warehouse for advanced analytics and historical reporting** - **Secure Authentication using Supabase Auth with email and password** - **Attendance Analytics with department-wise reports and attendance percentage tracking** ### Tech stack: Python ### Why it matters: The Facial Recognition Based Attendance System offers a comprehensive, enterprise-grade solution for efficient attendance management in educational institutions. Its advanced facial recognition technology, combined with real-time tracking and comprehensive reporting, streamlines the attendance process, reduces manual errors, and provides valuable analytics for educational institutions.

### Overview: This comprehensive survey explores the advancements and challenges in Vision-Language Models (VLMs), including fine-tuning, pre-trained models, prompt engineering, adapters, and benchmark datasets. It aims to provide a foundational resource for researchers and practitioners seeking to optimize VLMs for diverse real-world scenarios. ### Key Features: - **Fine-Tuning**: Techniques for adapting pre-trained models to specific tasks. - **Pre-trained Models**: Comprehensive overview of pre-trained models and their applications. - **Prompt Engineering**: Strategies for using prompts to enhance model performance. - **Adapters**: Modules that adapt pre-trained models for specific tasks. - **Benchmark Datasets**: Evaluation metrics and datasets used to benchmark VLM performance. ### Tech Stack: Python, Markdown, and GitHub. ### Why It Matters: This survey provides a critical analysis of the latest trends, challenges, and future research directions in VLM optimization, making it essential for researchers and practitioners aiming to improve VLM performance in diverse real-world applications. ### Repository: SufyanDanish/VLM-Survey- ### GitHub Stars: 17 ### Primary Language: Unknown ### Open Challenges: The survey highlights the need for scalable, generalizable, and bias-aware VLM designs, as well as advancements in symbolic reasoning, multilingual adaptation, and energy-efficient VLM design.

### Overview: StealthIt is an invisible AI desktop assistant designed for live meeting notes, technical interviews, and more. It uses state-of-the-art AI models like Claude, OpenAI, and Ollama, while remaining completely hidden from screen captures and video calls. ### Key Features: - **True Stealth Invisibility**: Excluded from the Windows Taskbar and 100% invisible to screen recording tools and video calls. - **Local Whisper Voice & Meeting Transcription**: Press `Ctrl+R` to record system audio or microphone. - **Instant Visual Analysis**: Provides real-time summaries and analysis of meetings, interviews, and technical documents. - **Live Coding & Presentation Prompting**: Floats over IDEs and platforms like LeetCode, enabling real-time problem-solving and presentation prompts. - **Voice Transcription**: Transcribes audio from Zoom, Teams, Google Meet, and other video calls. - **Premium LLM Support**: Compatible with Anthropic, OpenAI, OpenRouter, Gemini, and local Ollama. ### Tech Stack: Python and PySide6/Qt. ### Why It Matters: StealthIt offers a powerful and invisible AI assistant for professionals, enhancing productivity and meeting efficiency without compromising on security or functionality.

### Overview: finestSAM is a configuration-driven framework for fine-tuning Meta AI's Segment Anything Model (SAM) on custom datasets in COCO format. It allows developers to control various details such as training strategy, prompt supervision, frozen components, optimization knobs, and dataset handling, making fine-tuning reproducible and highly controllable. ### Key features: - **Configuration-driven**: Fine-tuning parameters can be easily configured via `finestSAM/config.py`. - **Training Strategy**: Supports classic fine-tuning or LoRA adapters (encoder and/or decoder). - **Prompt Supervision**: Options for boxes, points, or masks for supervision. - **SAM Components**: Control over which components of SAM are frozen. - **Optimization**: Toggles for losses, metrics, schedulers, precision, and device selection. - **Data Handling**: Options for dataset split/caching and dataset-driven normalization statistics. ### Tech stack: Lightning AI Fabric ### Why it matters: finestSAM simplifies the fine-tuning process by making it reproducible and highly controllable, allowing developers to fine-tune the model according to their specific needs. This framework is designed to be scalable, compatible with SAM inference, and supports automatic prediction workflows, making it an essential tool for researchers and practitioners working with SAM models.

Overview: This repository provides Node.js examples to help developers get started with the Amazon Bedrock service. Key features: - Amazon Bedrock Service Integration - Node.js Examples - Simple and Practical Solutions Tech stack: Node.js Why it matters: This repository offers a straightforward way for developers to begin working with Amazon Bedrock, making it easier to integrate and utilize the service in their projects.

### Overview: A curated list of awesome Pinecone resources, libraries, tools, and applications. ### Key features: - Quickstart guide for setting up and using Pinecone Database - Sample notebooks and code walkthroughs for various Pinecone applications - Explanation of Pinecone's database architecture - Release notes and full-text search capabilities - Guide on building a RAG chatbot without coding - Populating Pinecone from a website using n8n - Introduction to Pinecone Nexus, the knowledge engine for agents - Announcement of the open-source PostgreSQL extension, pg_pinecone - Launch post for Pinecone Nexus and an overview of the knowledge engine for agents - Launch post for Full-Text Search and keyword search capabilities in Pinecone - Hands-on walkthroughs for indexing and querying with Pinecone in Python ### Tech stack: Python, JavaScript, PostgreSQL, and various open-source tools and libraries. ### Why it matters: Pinecone is a powerful vector database that enables developers to easily integrate vector search capabilities into their applications. The curated list of resources and tools provided in this repository makes it easy for developers to get started with Pinecone and build intelligent applications that can understand and respond to natural language queries.

Overview: This repository contains a fine-tuned version of the unsloth/DeepSeek-R1-Distill-Llama-8B model, named abhi9ab/DeepSeek-R1-Distill-Llama-8B-finance-v1, optimized for financial reasoning tasks. Key features: - Fine-tuned with LoRA on a reduced subset of the Finance-Instruct-500k dataset - Utilizes the base model from unsloth/DeepSeek-R1-Distill-Llama-8B - Platform: Free-tier Kaggle Notebook - Libraries: Hugging Face Transformers, Unsloth, Weights and Biases (wandb), PyTorch Tech stack: Python, Hugging Face Transformers, Unsloth, Weights and Biases (wandb), PyTorch Why it matters: This model enhances the base model's performance on financial tasks by fine-tuning it on a specialized dataset, leveraging low-rank adaptation for efficient training.

Overview: afdEP-comm is an expert-parallel communication system for Attention–FFN Disaggregation (AFD) built on DeepEP V2, designed to optimize performance and reduce latency. Key features: - Expert-parallel communication between Attention GPUs and FFN GPUs. - Vendor of DeepEP V2's kernels and specialized them for the two roles. - Routing, quantization, expert compute, and serving scheduling remain with the caller. - Benchmarks run against FastAFD's DeepEP M2N, showcasing improvements in performance and efficiency. Tech stack: Python Why it matters: By optimizing the communication between different components of the AFD system, afdEP-comm significantly reduces latency and improves overall performance, making it a valuable tool for developers working on high-performance computing and machine learning applications.

### Overview: This project aims to fine-tune a PyTorch object detector for the purpose of detecting road damage in the RDD2022 dataset, with the goal of participating in the Crowdsensing-based Road Damage Detection Challenge (CRDDC2022). ### Key features: - Fine-tuning of a ResNet-50 FPN backbone with pre-trained weights for road crack detection. - Implementation of a PyTorch object detector for road damage detection. - Training on the Norway training subset of the RDD2022 dataset. - Evaluation on the test dataset of the challenge with results of 0.18, 0.21, and 0.31 for 1, 2, and 3 trainable backbone layers respectively. ### Tech stack: Python ### Why it matters: This project demonstrates the effectiveness of fine-tuning pre-trained models for specific tasks, particularly in detecting subtle road damage in challenging scenarios.

### Overview: Vortex is a self-hosted RAG application that allows users to chat with their documents using any LLM provider. It features a clean, modern interface, hybrid search, re-ranking, source citations, and multi-provider support. ### Key Features: - **Hybrid Search**: Combines vector similarity with keyword matching for better retrieval. - **Re-ranking**: Re-ranks results using exact phrase matching, keyword density, and position scoring. - **Source Citations**: Responses include `[n]` citation notation linking to specific document sources. - **Semantic Chunking**: Documents are split into 1500-character chunks with 300-character overlap for better context preservation. - **Multi-Provider Support**: Supports 5 LLM providers (OpenAI, Anthropic, Google, xAI, OpenRouter) and 4 embedding providers (Xenova/Transformers.js, OpenAI, Google, OpenRouter). - **Knowledge Base Management**: Multiple knowledge bases, each with its own embedding model, and a dashboard showing document counts, conversation counts, and model badges. - **Document Ingestion**: Uploads PDFs and text files, or ingests content from any URL, with automatic chunking and embedding. - **Streaming Chat**: Real-time streaming responses with conversation persistence and full chat history per knowledge base. - **Settings Page**: Configures LLM provider, model, temperature, API keys, and default embedding model, with encrypted API keys and session refresh on every request. - **Security**: Auth middleware with Supabase SSR cookie-based auth,

### Overview: AiiStream Q3.6 is a state-of-the-art expert-streaming engine that runs a 35-billion-parameter model on a 16 GB Mac, achieving byte-identical output without retraining or reduced routing. ### Key features: - **Byte-identical Output**: Generated tokens are identical to the unmodified 4-bit model. - **No Retraining**: No need for retraining the model. - **No Reduced Routing**: No expert skipping, top-k reduction, or substitute experts. - **Runs on Apple Silicon**: Compatible with Apple M5 and 16 GB unified memory. - **High Decode Throughput**: Measured at 15.0 tok/s on an 8 GB Mac, with a peak of 1.7 GB of MLX memory. - **Client Disconnect**: Generation stops after the client disconnects, unaffected by subsequent requests. - **Memory Efficiency**: Stable memory usage, with peak memory of 1.7 GB during generation. - **Output Consistency**: Byte-identical to the stock MLX model, including tool calls and seeded sampling. ### Tech stack: Python ### Why it matters: AiiStream Q3.6 is an innovative solution for expert-streaming, offering high decode throughput and consistent output without the need for retraining or reduced routing. It is designed for use on Apple Silicon and runs on a 16 GB Mac, making it ideal for developers and researchers looking to improve their streaming models.