Projects

Overview: This project is a canary release analyzer for MLOps platforms, designed to ensure model serving decisions are documented and reviewed before traffic expansion based on error rate, tail latency, and drift. Key features: - Canary latency, error rate, and drift checks - Decision rules for severe and moderate drift, p95 latency, and promotion - JSON output showing matched reasons - CLI implementation in Python Tech stack: Python Why it matters: Ensures clear documentation and informed decision-making in model serving rollouts, preventing traffic expansion based on unreliable metrics. Repository: Klokesh493/model-serving-canary-analyzer Primary language: Python GitHub stars: 0 GitHub description: Model-serving canary release analyzer for MLOps platforms.

### Overview: This project is an end-to-end MLOps capstone developed using Databricks, PySpark, Scikit-learn, MLflow, Unity Catalog, and model serving. It demonstrates the entire ML lifecycle, from data exploration to model deployment. ### Key features: - Data exploration and preprocessing using Python and Spark. - Training and evaluating multiple classification algorithms. - Experiment tracking with MLflow for parameter and metric logging. - Model selection based on accuracy, F1 score, and ROC-AUC. - Logging the selected model with input/output signatures for model management. - Registering the model in Databricks Unity Catalog. - Deploying a real-time serving endpoint for model inference. ### Tech stack: Python, Databricks Free Edition, Apache Spark / PySpark, Pandas, NumPy, Scikit-learn, MLflow, Unity Catalog, Databricks Model Serving, Git and GitHub. ### Why it matters: This project serves as an educational demonstration of an end-to-end ML lifecycle, showcasing how to handle structured datasets, train models, track experiments, compare models, and deploy models in a production-ready environment.

### Overview: This project is a FastAPI-based machine learning model serving application built with Docker, Docker Compose, and Google Cloud Run. The API exposes prediction, health check, and model information endpoints, configured for deployment on Google Cloud Run. ### Key Features: - **FastAPI** API for serving the model. - **Docker** for containerization. - **Docker Compose** to orchestrate the application. - **Google Cloud Run** for deployment. - **Pydantic** for data validation. - **Logistic Regression** model for demonstration. - **Model information** exposed via endpoints. ### Tech Stack: - Python ### Why it Matters: This project showcases how to build and deploy a machine learning model serving application using FastAPI, Docker, and Google Cloud Run. It demonstrates the integration of ML model serving with backend infrastructure, making it accessible and scalable. ### Repository: KornHolio216/fastapi-ml-api - **Primary Language**: Python - **GitHub Stars**: 0 - **Description**: ML model serving project built with FastAPI and Docker. The API exposes prediction, health check and model info endpoints, with configuration prepared for deployment on Google Cloud Run.

### Overview: This project demonstrates an end-to-end machine learning pipeline that includes training, FastAPI serving, and drift monitoring. It uses a dataset-agnostic approach, featuring a built-in scikit-learn dataset and sample-size floor derived from measured false-alarm rates. ### Key features: - **Training**: Experiment tracking and MLflow integration. - **Serving**: FastAPI endpoint for serving the trained model. - **Drift Monitoring**: PSI and KS drift detection with sample-size floor for drift detection. ### Tech stack: Python ### Why it matters: This project showcases best practices in production machine learning operations (MLOps), including dataset-agnostic training, CI/CD pipeline, and drift monitoring with a sample-size floor derived from empirical data.

Overview: This repository is a comprehensive FastAPI learning resource that covers REST APIs, Pydantic, HTTP methods, Machine Learning model serving, Docker containerization, and AWS deployment. It includes practical code examples and hands-on projects, making it ideal for developers looking to build production-oriented APIs for Machine Learning models. Key features: - FastAPI for Machine Learning - Comprehensive documentation and hands-on projects - Practical code examples - Dockerization and AWS deployment Tech stack: The repository uses Jupyter Notebook, FastAPI, Pydantic, Docker, and AWS. Why it matters: This repository provides a structured learning path from the basics of FastAPI and REST APIs to advanced topics like serving Machine Learning models and deploying on AWS, making it a valuable resource for developers looking to build robust and scalable APIs for Machine Learning applications.

### Overview: This project is an image anti-spoofing experiment repository, featuring a comparison of CNN and ViT models. It includes a training notebook, PyTorch architecture definitions, image preprocessing, an inference API, and a browser upload interface for real and fake image classification. ### Key features: - **Model Comparison**: A comparison between CNN and ViT models for image classification. - **Training Notebook**: Contains a training notebook for model training. - **Inference API**: An API that loads a pre-trained model from Hugging Face Hub and serves predictions. - **Image Preprocessing**: Includes image resizing and normalization for input preprocessing. - **Dataset & Method**: Uses a merged dataset of original and fake images, with a baseline CNN, deeper CNN, and ViT model for classification. ### Tech stack: Python, PyTorch, FastAPI, React, Hugging Face Hub ### Why it matters: This project provides a practical demonstration of image classification using CNN and ViT models, showcasing their effectiveness in distinguishing between real and fake images. The use of Hugging Face Hub for model artifacts ensures reproducibility and accessibility.

### Overview: This project is a React and FastAPI application designed to classify speech as either fake or real. It utilizes a PyTorch CNN model trained on Mel spectrograms, serving predictions through a deployed prediction interface. The application includes a frontend hosted on Netlify and a separate training repository for the model artifacts. ### Key Features: - **Real-time Prediction:** Classifies uploaded speech as fake or real. - **Model Artifacts:** Uses a fine-tuned baseline CNN model from Hugging Face Hub. - **Backend Deployment:** Runs inference using a CPU-compatible PyTorch model. - **Frontend Hosting:** Hosted on Netlify with a backend configuration for improved performance. ### Tech Stack: - **Inference:** Python, PyTorch, Torchaudio, Hugging Face Hub. - **API:** FastAPI, Uvicorn. - **UI:** React 18, Create React App. - **Hosting:** Netlify frontend, Render backend configuration. ### Why it Matters: This project provides a practical solution for detecting audio spoofing, which is crucial in various applications such as security systems and audio verification. The use of advanced machine learning techniques ensures high accuracy, making it valuable for real-world applications. ### Repository: - **GitHub Repository:** Laabh-Gupta/Voice-Anti-Spoofing-Web-App - **GitHub Stars:** 0 - **GitHub Description:** React + FastAPI speech anti-spoofing app serving a PyTorch CNN on Mel spectrograms. Model artifacts on Hugging Face; separate training repository.

Overview: SmarterRouter is an intelligent gateway and VRAM-aware router designed for Ollama, llama.cpp, and OpenAI, featuring semantic caching, model profiling, and automatic failover for local AI labs. Key features: - Semantic caching - Model profiling - Automatic failover Tech stack: Python Why it matters: SmarterRouter optimizes performance and reliability for AI research and development, ensuring seamless integration and efficient resource management.

Overview: This project provides a Docker-based server for serving the OrcaBonsai Ternary Bonsai 2 27B model, ablated with the PrismML adapter, specifically targeting CUDA GPUs. The model is served without modifications, ensuring byte-identical output and runtime ablation strength. Key features: - No weights or binaries are modified, ensuring the model's integrity. - The model is served using the OrcaRouter runtime ablation, providing ablation strength as a runtime dial. - The model is served on a GPU, utilizing the latest CUDA 13.1.1 toolkit. - The model is served via a Docker container, making it easy to deploy and run. Tech stack: The project uses Docker and Compose for deployment, leveraging the latest CUDA toolkit for optimal performance. Why it matters: This project is crucial for developers and researchers who need to quickly deploy and test ablated models on CUDA GPUs, ensuring that the ablation effects are accurately reflected in the output without the need for re-quantization or modification of the original model weights.

Overview: This project is a REST API designed for real-time customer segmentation and ML-based scoring using FastAPI, Scikit-learn, SQLAlchemy, and Redis. It integrates customer data from various sources and provides machine learning-based insights into customer behavior and segmentation. Key features: - Real-time customer segmentation - Predictive analytics - Automatic API documentation Tech stack: Python 3.x, FastAPI, SQLAlchemy, Scikit-learn, Redis Why it matters: This API is crucial for businesses looking to gain real-time insights into customer behavior and segment their customer base effectively, improving marketing strategies and customer service.

Overview: This project provides a REST API for HR analytics with machine learning integration, offering standardized endpoints for personal analyses and predictive models. Key features: - Predictive models - Performance tracking - Automated reports Tech stack: Python 3.x, FastAPI, TensorFlow, PostgreSQL, Redis Why it matters: This project leverages modern technologies like FastAPI for efficient API development, TensorFlow for machine learning, and PostgreSQL and Redis for robust data management and caching, making it suitable for real-time analytics and predictive modeling in HR. Repository: Lendrx/data-connector-hr Primary language: Jupyter Notebook GitHub stars: 0 GitHub description: REST API zur Analyse von Personaldaten und Echtzeit-Vorhersagen mit FastAPI, TensorFlow, PostgreSQL und Redis Caching.

### Overview: This project is a REST-API for training and inference of an Iris classifier using FastAPI and Docker. It provides a simple and scalable solution for machine learning tasks. ### Key Features: - **Training and Inference**: The API supports both training and inference of the Iris model. - **Docker Integration**: The project is built with Docker to ensure easy deployment and scalability. - **FastAPI**: Utilizes FastAPI for efficient and clean API development. ### Tech Stack: - Python ### Why It Matters: This project is ideal for developers looking to quickly set up a machine learning API with Docker and FastAPI, making it easy to integrate into existing projects or as a standalone service.

### Overview: This project focuses on implementing GPU-accelerated vector search algorithms using Triton, a framework for GPU kernels. The primary goal is to benchmark the performance of these kernels against traditional NumPy implementations, demonstrating significant speedup across various distance metrics and search techniques. ### Key Features: - **Vector Search Algorithms**: Implementations of KNN, K-means, and IVF-style approximate nearest-neighbour search using Triton kernels. - **Distance Metrics**: Support for L2, cosine, dot product, and Manhattan distances. - **Request Queue and Dynamic Batching**: A RAG service with a request queue and dynamic batching mechanism, optimized for handling large datasets and varying input sizes. ### Tech Stack: - Python ### Why It Matters: The project showcases the effectiveness of GPU-accelerated kernels in improving the performance of vector search algorithms, particularly in terms of speedup and scalability. By benchmarking against NumPy, it highlights the significant performance gains achievable with Triton, making it a valuable tool for researchers and developers working on large-scale vector search applications.

Overview: This project is about the full ML engineering loop around a stock price prediction model: data preprocessing, feature engineering, model training, evaluation, experiment tracking, and serving the trained model as a FastAPI endpoint. Key features: - Data preprocessing and feature engineering with Pandas - Model training and evaluation with Scikit-learn - Experiment tracking (parameters, metrics, model versions) via MLflow - FastAPI endpoint serving the trained model Tech stack: Python, Scikit-learn, MLflow, FastAPI, Pandas Why it matters: This project demonstrates the full ML engineering loop and the separation of training pipeline and serving layer, which is applicable to any ML model, regardless of its prediction accuracy.

Overview: SmarterRouter is an intelligent LLM gateway and VRAM-aware router designed for Ollama, llama.cpp, and OpenAI. It includes semantic caching, model profiling, and automatic failover for local AI labs. Key features: - Semantic caching - Model profiling - Automatic failover Tech stack: Python Why it matters: SmarterRouter enhances the performance and reliability of AI models by optimizing resource usage and ensuring smooth operation in local AI labs.

Overview: This project is a production-style machine learning API for credit risk scoring, utilizing the UCI Statlog German Credit Data set. It exports a versioned model package, optimizes cost-sensitive threshold, and serves predictions through a FastAPI application packaged for Docker deployment. Key features: - Training multiple classification models (Logistic Regression, Random Forest, XGBoost) - Applying cost-sensitive threshold optimization - Exporting a versioned model package - Simulating inference from raw applicant JSON - Building a FastAPI service - Adding automated tests - Preparing Docker deployment Tech stack: FastAPI, Docker, Jupyter Notebook Why it matters: This project demonstrates the entire engineering workflow needed to turn a trained model into a reproducible inference service, including model training, cost-sensitive threshold optimization, and Docker deployment for production-ready application.

### Overview: This project is a FastAPI wrapper for quick ML model testing. It provides pre-made endpoints for registration, loading, prediction, and metrics, along with authentication and Prometheus metrics and an inference log in PostgreSQL. ### Key features: - Pre-made endpoints for registration, loading, prediction, and metrics - FastAPI wrapper for easy model testing - Real-time metrics and inference logs in PostgreSQL - Built-in authentication and Prometheus metrics ### Tech stack: Python ### Why it matters: This project streamlines the process of testing and deploying ML models, making it easier for developers to quickly integrate models into their applications. The pre-made endpoints, real-time metrics, and built-in authentication ensure that the models are always ready and easily accessible. ### Repository: MeJustBear/python-ml-service ### Primary language: Python ### GitHub stars: 0 ### GitHub description: FastAPI wrapper for quick ML model testing: register load, predict and metrics functions — get ready-made endpoints, authentication, Prometheus metrics and an inference log in PostgreSQL ### README excerpt: # mlwrap FastAPI-обёртка для быстрых тестов ML-моделей. Модель подключается тремя функциями — загрузка, инференс, сбор метрик, — которые регистрируются в реестре по имени модели. Эндпоинты, аутентификация, метрики Prometheus и журнал запусков в БД уже готовы. **Справка по

Overview: LinML is an open-source project aimed at providing a Linux counterpart to Windows ML (WinML), offering a unified, hardware-abstracted inference layer, CLI, and catalog library for ONNX Runtime. Key features: - LinML provides a Linux-first inference integration layer, CLI, and hardware provider catalog. - LinML.Catalog automates the vendoring, SHA-256 verification, dynamic registration, and session attachment for external plugin execution providers. - LinML supports ONNX Runtime with versions 1.31.x. Tech stack: C# Why it matters: LinML aims to democratize machine learning on Linux by providing a unified and hardware-agnostic inference layer, making it accessible to developers on Linux environments.

Overview: LinML is an unofficial Linux counterpart to Windows ML (WinML), providing a Linux-first ONNX Runtime inference layer, CLI, and hardware provider catalog. Key features: - LinML CLI - LinML.Catalog library - LinML.Catalog library for vendorizing, SHA-256 verification, and dynamic registration of external plugin execution providers Tech stack: C# Why it matters: LinML aims to bring the ergonomic principles of WinML to Linux environments, offering honest execution provider discovery, explicit selection policies, reproducible inference execution, and diagnosable fallback across various hardware providers.

Overview: SmarterRouter is an advanced gateway and VRAM-aware router designed for Ollama, llama.cpp, and OpenAI, featuring semantic caching, model profiling, and automatic failover for local AI labs. Key features: - Semantic Caching - Model Profiling - Automatic Failover Tech stack: Python Why it matters: SmarterRouter enhances the performance and reliability of AI models by intelligently managing resource usage and ensuring optimal operation, making it crucial for efficient AI development and deployment.