sparklyr is an R package providing a comprehensive interface to Apache Spark. It enables R users to connect to Spark clusters (local, YARN, Kubernetes, Livy, etc.), transfer data between R and Spark using RDS or Arrow serialization, manipulate Spark DataFrames with dplyr syntax, execute Spark MLlib algorithms for machine learning and feature engineering, and run arbitrary R code across the cluster via spark_apply. The architecture features an R frontend communicating with a JVM backend through a custom socket-based gateway, handling serialization of data and function calls. Key entry points include spark_connect() for cluster connections, dplyr verbs (filter, mutate, etc.) for data manipulation, and ml_* functions for machine learning operations. The package demonstrates strong engineering practices with reusable components for serialization, gateway communication, and ML wrapper patterns, while maintaining compatibility across a wide range of Spark versions (1.5.2 to 4.0.x).
Docling is a production-grade Python library for converting diverse document formats (PDF, DOCX, PPTX, XLSX, HTML, LaTeX, EPUB, images, audio, video, email, EBCDIC, and more) into a unified, structured `DoclingDocument` representation. It is designed to serve as the document understanding layer for AI workflows, particularly Retrieval-Augmented Generation (RAG), agentic AI, and document extraction pipelines. The system provides a rich Python SDK, a Typer-based CLI (`docling convert`), a service client for remote conversion via `docling-serve`, and an MCP server. It supports multiple output formats including Markdown, JSON, HTML, DocTags, and plain text. Architecturally, Docling employs a modular pipeline system with pluggable backends for each input format and a stage-based processing model. The core pipeline (`StandardPdfPipeline`, `VlmPipeline`, `AsrPipeline`) orchestrates page-level operations (preprocessing, OCR, layout detection, table structure recognition, picture classification, reading order) and document-level enrichments (code/formula extraction, heading hierarchy, picture description). Inference engines are abstracted behind factory patterns supporting ONNX Runtime, HuggingFace Transformers, vLLM, MLX, KServe v2, and remote APIs. The library integrates with over 20 OCR engines and vision-language models, and includes a plugin system for extensibility. Key entry points include the `DocumentConverter` class, the CLI (`docling.cli.main`), and the service client (`docling.service_client`). External dependencies span PDF libraries (pypdfium2, docling-parse), ML frameworks (torch, transformers, onnxruntime), OCR engines (Tesseract, EasyOCR, RapidOCR), and cloud services (KServe, Azure, S3). The codebase is actively maintained with frequent releases, extensive test coverage, and a clear migration path from legacy components.
Koharu is a desktop application for manga and comic translation, built with a Rust backend (Tauri) and a React/TypeScript frontend. It orchestrates a multi-stage ML pipeline: page import, layout detection (YOLO, RT-DETR), speech bubble segmentation, OCR (PP-OCR, Manga OCR), translation (local LLMs via llama.cpp or cloud APIs), and inpainting (LaMa, FLUX.2, AOT-GAN). The system includes a custom deep learning framework (`koharu-torch`) with inference-only ports of common vision models, a scene graph for versioned project state, and a WebGPU/WASM canvas for interactive editing. The application supports multiple hardware backends (CUDA, ROCm, Vulkan, Metal) and can run entirely offline. Architecturally, the project is a monorepo with over 30 Rust crates and two JavaScript packages. The backend is organized around clearly separated concerns: model inference (`koharu-ml`), pipeline orchestration (`koharu-pipeline`), project state (`koharu-scene`), persistent storage (`koharu-storage`), and rendering (`koharu-renderer`/`koharu-rasterizer`). The frontend provides an editor with panels for page management, canvas, inspection, and an AI agent panel that uses OpenAI's Codex protocol for tool-driven project editing. The translation subsystem abstracts over local and remote providers with a common prompt and schema interface. While the primary purpose is an application, the ML model implementations are notable for their fidelity to upstream Python references, including exact preprocessing (LetterBox padding, pixel normalization) and postprocessing (NMS, mask decoding). This makes the repository valuable as a reference for Rust ports of computer vision models. However, the codebase is tightly coupled to specific model architectures and pinned revisions, limiting direct reuse without significant adaptation. The UI component library (`packages/ui`) and several backend abstractions (e.g., configuration, storage format, runtime dependency graph) are designed for reusability and could be extracted with minimal effort.
PlumeLens (鉴翎) is a local bird photo analysis and culling workstation for bird photographers. It uses on-device ML models—YOLOv26l for detection, a custom 11-keypoint pose estimator, CLIPIQA+/HyperIQA for quality assessment, and DINOv3 with LoRA adapters for species classification—to process photos without any cloud upload. The system scans user photo directories (read-only), performs detection, pose estimation, quality scoring, and species identification, then presents a grading interface (reject/record/usable/select) with manual review and species override capabilities. Export supports JPG/RAW with XMP sidecars, and a geographic archive module visualizes species sightings on an interactive map with conservation status. The architecture is a desktop application split into an Electron+React frontend and a Python FastAPI backend. The backend runs as a subprocess managed by Electron, communicating via REST APIs and Server-Sent Events for progress. The ML pipeline is orchestrated in `engine/pipeline/manager.py` with ONNX Runtime and PyTorch inferences. The frontend uses React 19, TanStack Query, Zustand, and a virtualized grid for handling large photo sets. The application is packaged for macOS (arm64) via electron-builder and PyInstaller. While the primary delivery is a desktop application, the repository also contains extensive model training scripts, evaluation frameworks (`evals/`), and dataset preparation utilities, making it a hybrid project with significant research and data engineering components. Model integrity is enforced via SHA-256 manifest verification at startup, and all inference is deterministic with versioned pipeline hashes to invalidate caches on updates.
Made With ML is an educational data science project that implements an end-to-end ML pipeline for classifying GitHub repositories into categories (NLP, computer vision, MLOps, other) based on their titles and descriptions. The core data lineage flows from raw CSV datasets (dataset.csv, holdout.csv) through text cleaning, tokenization via SciBERT, and stratified splitting into train/val/test sets, followed by distributed training with PyTorch and Ray Train, hyperparameter tuning via HyperOpt, evaluation with slice-based metrics using Snorkel, and serving through a FastAPI/Ray Serve deployment. The project demonstrates MLOps patterns including MLflow experiment tracking, data validation with Great Expectations, behavioral model testing, and CI/CD via GitHub Actions with Anyscale for cloud execution. The codebase is well-structured with clear separation of concerns across data.py, train.py, tune.py, evaluate.py, predict.py, serve.py, models.py, and utils.py. It includes comprehensive tests at the code, data, and model levels. However, the project shows signs of staleness with pinned dependencies from 2023 (Ray 2.7.0, transformers 4.28.1, torch 2.0.0) that have known security vulnerabilities and compatibility issues with current Python versions. The data pipeline components—particularly the CustomPreprocessor, stratified splitting, and text cleaning—are genuinely reusable patterns. The evaluation framework with slice-based metrics and behavioral tests (invariance, directional expectations, minimum functionality) represents solid ML engineering practice. The deployment configuration is tightly coupled to Anyscale's proprietary platform, limiting portability.
Mage is an open-source data pipeline orchestration platform that combines a visual notebook-style editor with a production-grade scheduling and execution engine. The system enables users to build, run, and monitor data pipelines composed of Python, SQL, R, and dbt blocks. Architecturally, it is a full-stack web application with a Python backend (Tornado-based server, SQLAlchemy ORM, REST API) and a Next.js/React frontend with a custom component library. Key entry points include the CLI (`mage` command), the web UI served by `mage_ai/server/server.py`, and the API layer at `mage_ai/api/` which exposes CRUD resources for pipelines, blocks, runs, schedules, users, and integrations. Data flows center around pipeline execution: blocks are connected in a DAG, each block can be a data loader, transformer, or exporter. Execution is handled by a configurable executor (local, K8s, EMR, PySpark) and scheduled via a built-in scheduler or external triggers. The platform integrates with dozens of data sources and destinations through a Singer-based integration framework, supports streaming pipelines, and includes advanced features like RBAC, version control, AI-assisted pipeline generation, and a command center for global search. It depends heavily on external services such as cloud storage (S3, GCS, Azure), databases (PostgreSQL, BigQuery, Snowflake, etc.), and authentication providers (OAuth, LDAP, OIDC).
TensorFlow Transform (tf.Transform) is a production-grade library for large-scale data preprocessing in TensorFlow pipelines. It computes global statistics (analyzers) over full datasets and applies them as transformations, ensuring train/serve consistency. The library integrates with Apache Beam for distributed execution and is part of the TFX ecosystem. Users define a preprocessing function using tft analyzers (e.g., min, max, mean, quantiles, vocabulary, TF-IDF, PCA) and mappers (e.g., scaling, bucketization, n-grams), which is traced and exported as a TensorFlow SavedModel. The architecture revolves around a graph of operation nodes (analyzer_nodes.py) that are traversed and translated into Beam PTransforms. Key features include incremental caching, combiner packing for performance, deep copy of Beam collections, and schema inference. The codebase is well-structured with extensive tests, but carries technical debt from TF1/TF2 migration, some known test failures (xfail), and experimental APIs. It is actively maintained within the TFX ecosystem.
ScienceBeam Parser is a scientific document parsing library and service that converts PDF/DOCX documents into structured formats (TEI/JATS XML) using machine learning models. It provides a FastAPI-based web service with GROBID-compatible endpoints, a CLI for training data generation, and a modular pipeline of models for segmentation, header, citation, reference extraction, etc. The system includes benchmarking tools for performance evaluation against GROBID baselines and training utilities for model improvement. While it contains research-oriented components (benchmarking, training), its primary character is as a deployable application/library for document processing.