This article outlines the essential concepts, data preparation, model creation, training, deployment, and verification steps for testing machine‑learning models, highlighting dataset requirements, algorithm categories, framework choices, resource considerations, and provides a sample inference request.
The article defines the AI architect role and outlines three essential capabilities—mastery of AI technologies and development workflow, deep business understanding with strong abstraction ability, and the design and implementation of efficient, scalable AI solutions—explaining why each is critical for successful AI product delivery.
This guide walks you through using Meta's open‑source Llama 2 models on Alibaba Cloud's PAI platform, covering low‑code LoRA fine‑tuning, full‑parameter fine‑tuning with PAI‑DSW, and rapid WebUI deployment via PAI‑EAS, complete with step‑by‑step instructions, code snippets, and resource requirements.
The article describes Sunshine Insurance Group's Zhèngyán Large Model Open Platform, detailing its three‑layer architecture, AutoTrain tool, self‑developed LLM, smart routing, plugin marketplace, intelligent review, and how these capabilities empower insurance marketing, sales, service, and management through AI‑driven solutions.
This article presents an end‑to‑end overview of Alibaba Cloud’s Machine Learning PAI platform, detailing the three‑stage ML workflow, challenges in model development, the role of pre‑trained and open‑source models, PAI’s architecture, a hands‑on demo, and MLOps best practices for efficient model deployment.
This article details the step‑by‑step process of preparing the environment, merging weights, installing dependencies, running inference, evaluating Vicuna‑7B against other models, and attempting fine‑tuning, while highlighting performance results, encountered issues, and future work for large language model deployment.
The article analyzes the operational, stability, and cost challenges of Haro’s AI platform, explains why a serverless FaaS architecture—specifically Knative—was selected, and details the implementation steps, performance gains, and future scenarios for AI workloads.
The team embedded a full machine‑learning ranking pipeline as an Elasticsearch plug‑in—combining real‑time and offline feature stores, hot‑loadable model jars via Dragonfly, an MLeap execution engine, and a DSL for feature definition—replacing the coarse‑ranking logistic‑regression with a tree model that adds ~10 ms latency but yields a 1.2 % AB‑test lift, while maintaining high throughput, low CPU usage, and supporting future batch deep‑learning rescoring.
Tencent’s Taiji machine learning platform, a cloud‑native, distributed parameter‑server system, provides end‑to‑end MLOps for advertising by integrating data ingestion, feature engineering, model training, evaluation, deployment, and monitoring, supporting massive models up to billions of parameters while improving efficiency, scalability, and resource management.
The Xianyu recommendation system, built by backend expert Wan Xiaoyong, evolved from offline scoring to a full‑graph, serverless recall‑ranking pipeline that tackles C2C uncertainties through centralized feature engineering, model compression, staged deployment, flexible experimentation, robust governance, and plans for automated attribution and interpretability.
This article reviews Ant Group's "thousand‑by‑thousand" model strategy for mobile AI, detailing the challenges of device compute and user heterogeneity, the engineering upgrades for offline development and online release, and the measurable business improvements achieved.
This article introduces the EasyModeling framework built on Alibaba Cloud's PAI platform, detailing its modular design, high reusability, integration with deep‑learning libraries, automated hyper‑parameter tuning, deployment scenarios, and a real‑world case study using RoBERTa for dish‑name standardization, demonstrating significant performance gains.
This article examines the challenges of deploying dozens of deep‑learning models at Zuoyebang and compares three serving architectures—Gunicorn + Flask + Transformers, Tornado + PyTorch, and Tornado + Triton—highlighting performance trade‑offs and presenting a final high‑concurrency, low‑latency solution in production.
This article outlines the background, standards, evaluation items, process, and registration details of a newly launched MLOps development management maturity assessment designed to accelerate AI model delivery and improve operational efficiency across teams.
vivo’s machine‑learning platform, built for its massive app‑store and e‑commerce ecosystem, streamlines data processing, model training, and deployment through quota‑based resource management, a custom ultra‑large‑scale TensorFlow‑vlps framework, OpenAPI‑driven training, and Jupyter‑integrated interactive development, boosting efficiency for billions of samples and features.
The paper proposes a full quantization pipeline for YOLOv6 that combines a re‑parameterization optimizer, partial PTQ, channel‑wise distillation, graph‑scale merging, and GPU‑offloaded preprocessing, enabling an INT8 model to retain ~42 % mAP while delivering over 200 % throughput increase and 40 % QPS gain versus FP16.
This article describes the production challenges of using BERT for large‑scale text classification at Zuoyebang, explores lightweight alternatives such as knowledge distillation, pruning and quantization, and details a teacher‑student‑active‑learning pipeline that trains a TextCNN model to match BERT performance while dramatically reducing GPU consumption and improving throughput.
Meituan/Dianping’s edge‑AI re‑ranking system moves large‑scale deep‑learning models onto users’ devices, using dense networks and cloud‑served embeddings, advanced feedback‑sequence and multi‑view attention models, and aggressive compression to deliver real‑time, privacy‑preserving search personalization that boosts click‑through rates by up to 0.43 %.
This article provides an in‑depth overview of the TensorFlow ecosystem, covering Keras modeling productivity tools, classic model examples, AutoKeras and KerasTuner for automated search, data preprocessing pipelines, performance profiling, model optimization, and multiple deployment strategies for servers, browsers, and edge devices.
This article introduces the open‑source WantWords reverse‑dictionary project, explains its token‑based processing pipeline, walks through Python implementation and model invocation with Hugging Face’s Transformers, reviews NLP’s historical evolution, and shows how front‑end developers can quickly integrate NLP models into products.
This article details the architecture, incremental model training, feature engineering, and deployment strategies that enabled NetEase Cloud Music to achieve real‑time live‑stream recommendation, covering business background, multi‑objective modeling, real‑time feature pipelines, sample attribution, feature admission, and online performance results.
This article compares PyTorch and TensorFlow in 2022 across model availability, deployment ease, and ecosystem support, using data from HuggingFace, research papers, and industry tools, and offers tailored recommendations for industry engineers, researchers, educators, career changers, hobbyists, and beginners.
The team replaced the cumbersome ABox deployment stack with Seldon‑based cloud‑native serving on Kubernetes, unifying TensorFlow and other framework models, adding GPU sharing, automated CRUD, per‑model ingress, monitoring, and log collection, achieving scalable, fault‑tolerant, zero‑downtime model deployment.
The article examines three key design dimensions of modern AI platforms—programming language choice, runtime environment isolation, and scheduling/resource management—while also discussing challenges in model deployment such as algorithm diversity, resource usage patterns, and architectural generality, proposing Kubernetes‑based solutions and Arrow‑based data sharing to achieve efficient, scalable AI services.
The article examines three key design dimensions of modern AI platforms—choice of programming language, runtime environment isolation, and model deployment—highlighting how Python’s dominance, container‑based resource management, and efficient data sharing shape platform architecture and performance.
Discover how Pipcook 2.0 dramatically speeds up machine‑learning workflows for web developers—cutting installation to under 20 seconds, enabling rapid model training, prediction, and deployment via concise JSON pipelines, with step‑by‑step guidance, code snippets, and practical examples for image and text classification.
WanFang upgraded its academic paper retrieval system by adopting PaddleNLP’s Chinese pre‑trained Sentence‑BERT models, using weakly supervised SimCSE data and Milvus vector indexing, compressing the transformer for TensorRT‑accelerated inference, achieving 70% better matching quality and 2600 QPS latency‑optimized performance.
This article explains how a high‑traffic map service captures road features using client‑side computer‑vision models, details the deployment of many CNNs with the lightweight MNN engine on memory‑constrained devices, and shares practical memory‑saving techniques, inference scheduling, and error‑analysis methods.
Meituan’s Turing OS unifies the end‑to‑end machine‑learning lifecycle—data preprocessing, feature generation, model training, deployment, online prediction and A/B testing—through a lightweight SDK, plugin‑based algorithms, DAG orchestration, sandbox validation and replay tools, cutting algorithm iteration from days to hours while handling billions of daily predictions.
This article demonstrates how to convert traditional scikit‑learn and hmmlearn machine‑learning models into ONNX format and integrate them into a C++ gRPC service for fast online inference, covering environment setup, model conversion, custom operators, performance testing, and end‑to‑end pipeline construction.
This article explains how to build a code clone detection service using a CodeBERT classification model, create a custom TorchServe handler, package the model with torch-model-archiver, launch the service, and test it with example code pairs to demonstrate clone and non‑clone predictions.
The AMAP‑TECH competition challenged participants to infer real‑time road conditions from in‑vehicle video, prompting the authors to combine lane‑wise vehicle detection with LightGBM and later an end‑to‑end DenseNet‑GRU model, augment data, ensemble five networks, and achieve a 0.7237 F1 score while outlining future deployment and research directions.
This lecture details the end‑to‑end AI architecture for unmanned stores, covering algorithm module selection, calibration, face recognition, multi‑task detection, tracking, recommendation, data collection, augmentation, model training, and GPU‑accelerated deployment to achieve real‑time performance and high accuracy.
This guide explains the core concepts of TensorFlow Serving, shows how to prepare Docker images, save TensorFlow 2.x models in various formats, configure version policies, warm‑up models, start the service, and invoke it via gRPC or HTTP with complete code examples.
This article explains the core concepts of TensorFlow Serving, shows how to prepare the environment with Docker, convert TensorFlow 2.x models to the SavedModel format, configure version policies, warm‑up the service, and invoke predictions via gRPC or HTTP interfaces.
This article explains how to train and deploy convolutional neural networks directly on iOS and Android devices using Core ML, NNAPI, and TensorFlow Lite, compares performance with desktop TensorFlow, and provides practical code snippets and build‑time tips for mobile AI development.
This article explains the background of on‑device AI, introduces the Mobile Machine Learning (MML) framework and its layered architecture, details the core utilities such as model decryption and task scheduling, and provides a step‑by‑step code guide for initializing, preprocessing, inference, post‑processing, and releasing resources on mobile platforms.
This article analyzes the challenges of managing the entire machine‑learning lifecycle, reviews existing traceability approaches, and proposes comprehensive methods for versioned management of model training, prediction, and online service to improve efficiency, reproducibility, and maintenance of AI systems.
This article explains how Alibaba's Frontend Intelligent Project automatically generated 79.34% of the Double‑11 page code by recognizing business modules from visual drafts using data‑augmented samples, traditional multi‑class machine‑learning models, and a pipeline of preprocessing, model training, deployment, and OOD handling.
This tutorial walks university students through the complete workflow of using the CIFAR‑10 dataset to train a convolutional neural network for image classification and then deploying the model as an online inference service on the UAI‑Train and UAI‑Inference platforms.
This article explains how to prepare the Docker environment, install TensorFlow Serving (CPU and GPU versions), convert a YOLO V3 checkpoint to SavedModel, deploy the model as a service, warm‑up and manage versions, invoke it via gRPC and HTTP, and compare CPU versus GPU inference performance.
This article describes the end‑to‑end development of a multi‑class NLP system for Ctrip airline ticket customer service, covering problem analysis, data preprocessing, sample balancing, model architecture (TextCNN and Bi‑GRU), training strategies, performance evaluation, and online customization to achieve high accuracy in intent recognition.
This article describes a systematic, standardized, and automated “4+N” recommendation framework that unifies features, samples, models, and pipelines to accelerate private‑domain marketing recommendations across multiple scenarios while improving accuracy, efficiency, and business impact.
This article presents a comprehensive workflow for group risk perception, covering business requirement mining, data acquisition and understanding, feature engineering, model training and evaluation, deployment, and practical user applications, with detailed objectives, methods, and deliverables for each stage.
This article introduces Kubeflow, explains the typical machine‑learning model lifecycle, outlines Kubeflow’s core components and Kubernetes advantages, provides detailed server and storage configuration, walks through ksonnet and Kubeflow installation steps, and shows how to verify deployments and access the Kubeflow UI.
This article describes a comprehensive, end‑to‑end AI visual algorithm platform that unifies data collection, annotation, model training, deployment, testing, quality evaluation, and service gateways, illustrating how such integration improves transparency, efficiency, and quality across use cases like background removal, face swapping, and clothing recommendation.
This article introduces the open‑source H2O platform, walks through installing the Python package, demonstrates data import, model building with GBM and AutoML, evaluates results, explains model deployment via POJO/MOJO, and discusses the visual Flow UI and broader implications of automated modeling.
This article introduces the open‑source H2O platform, explains how to install and use its Python API for data loading, preprocessing, model training with GBM and AutoML, evaluates results with AUC, and describes model deployment via POJO/MOJO as well as the visual Flow UI, concluding with reflections on the role of automated modeling in data science.
This article offers a comprehensive, system‑engineer‑focused guide to Learning to Rank, covering fundamental machine‑learning concepts, evaluation metrics, training approaches, and a detailed online ranking architecture with feature, recall, and model governance, illustrated by real‑world examples from Meituan‑Dianping.
Meituan’s user‑growth team built a Wide‑Deep CTR prediction model, trained offline with Spark‑generated TFRecords, and deployed it via TensorFlow Serving on YARN, then applied request‑side multithreading, offline one‑hot preprocessing, XLA JIT compilation, and dedicated loading threads to cut end‑to‑end latency from ~18 ms to ~6 ms and eliminate model‑switch spikes.
These 43 practical rules, adapted from Martin Zinkevich’s “Rules of ML,” guide engineers through terminology, pipeline design, feature engineering, monitoring, and model deployment, offering actionable advice to avoid common pitfalls and build reliable, scalable machine‑learning‑driven products.
This article describes Getui's end‑to‑end machine‑learning data platform, covering business use cases, the full ML workflow from data ingestion and feature engineering to model training, deployment, monitoring, and the practical tools and solutions adopted to address common challenges in large‑scale AI projects.
This article provides a step‑by‑step tutorial on deploying trained deep‑learning models to production, covering client‑server architecture, load balancing with Nginx, using Gunicorn and Flask, cloud platform choices, autoscaling, CI/CD pipelines, and additional tools such as TensorFlow Serving and Docker.
Intelligent Titanium TI-ONE is a one‑stop ML IDE on Tencent Cloud offering integrated data preparation, drag‑and‑drop algorithm development, automatic hyperparameter tuning, multi‑level collaboration, one‑click model deployment, and support for major frameworks such as TensorFlow, PyTorch, Angel and XGBoost, plus commercial features via GaiaStack.
This guide walks you through training an Inception‑V3 model on the FER2013 dataset with TensorFlow 1.1, packaging the model, and deploying a scalable facial expression recognition API using UCloud's AI‑as‑a‑Service platform, including performance testing against GPU benchmarks.
The article describes how a three‑stage algorithm platform was designed and implemented to automate model deployment, unify feature processing, and provide service‑oriented model evaluation, debugging, and monitoring for machine‑learning applications in a large e‑commerce environment.
The article describes how Ctrip's algorithm platform was built in three stages to address deployment friction, reusable feature engineering, and model training, detailing the technical problems, Java/Python integration, code interfaces, transform configurations, and the eventual service‑oriented architecture.
This article shares practical engineering experiences for deploying machine learning models in various scenarios—real‑time low‑volume predictions via Rserve or Python‑httpserve, high‑throughput real‑time serving with PMML‑wrapped Java classes, and offline batch predictions using simple shell scripts—detailing tools, performance considerations, and implementation steps.
This article shares practical engineering experiences for deploying machine learning models in production, covering three typical scenarios—real‑time small data, real‑time large data, and offline predictions—and detailing how to use PMML, Rserve, Spark, shell scripts, and related tools to meet performance and operational requirements.
This article details how telecom operators can identify fraudulent ModemPOOL (cat‑pool) terminals and predict churn using data‑driven clustering and day‑interval warning models, covering metric selection, data exploration, k‑means clustering, model deployment, and performance evaluation.
