Operationalization of Machine Learning with Serverless Architecture: An Industrial Operationalization of Machine Learning with Serverless Architecture: An Industrial Implementation for Harmonized System Code Prediction

This paper presents a serverless MLOps framework orchestrating the complete ML lifecycle from data ingestion, training, deployment, monitoring, and retraining to using event-driven pipelines and managed services. The architecture is model-agnostic, supporting diverse inference patterns through standardized interfaces, enabling rapid adaptation without infrastructure overhead. We demonstrate practical applicability through an industrial implementation for Harmonized System (HS) code prediction, a compliance-critical task where short, unstructured product descriptions are mapped to standardized codes used by customs authorities in global trade. Frequent updates and ambiguous descriptions make classification challenging, with errors causing shipment delays and financial losses. Our solution uses a custom text embedding encoder and multiple deep learning architectures, with Text-CNN achieving 98 percent accuracy on ground truth data. Beyond accuracy, the pipeline ensures reproducibility, auditability, and SLA adherence under variable loads via auto-scaling. A key feature is automated A/B testing, enabling dynamic model selection and safe promotion in production. Cost-efficiency drives model choice; while transformers may achieve similar accuracy, their long-term operational costs are significantly higher. Deterministic classification with predictable latency and explainability is prioritized, though the architecture remains extensible to transformer variants and LLM-based inference. The paper first introduces the deep learning architectures with simulations and model comparisons, then discusses industrialization through serverless architecture, demonstrating automated retraining, prediction, and validation of HS codes. This work provides a replicable blueprint for operationalizing ML using serverless architecture, enabling enterprises to scale while optimizing performance and economics.

💡 Research Summary

The paper presents a comprehensive, server‑less MLOps framework designed to automate the classification of product descriptions into Harmonized System (HS) codes, a critical compliance task in global trade. The authors begin by outlining the business problem: thousands of products must be assigned six‑digit HS codes, and manual assignment is error‑prone, costly, and vulnerable to regulatory changes. Existing academic work largely relies on combined text‑and‑image models, ontology‑based approaches, or classical classifiers such as SVM and Random Forest, none of which fully address the scale, domain‑specific language, and operational constraints of an enterprise environment.

A high‑quality dataset is assembled from internal Schneider Electric records, comprising short and medium‑length textual descriptions together with verified HS codes. Labels are filtered by an internal “Assurance Level” system; only levels 3 and 4 (high confidence) are used for training, ensuring that the model learns from stable, expert‑validated data. The authors perform extensive preprocessing—lower‑casing, stop‑word removal, tokenization—and construct a custom embedding layer within the neural networks because generic Word2Vec/GloVe embeddings fail to capture the specialized engineering terminology.

Class imbalance is a major challenge because shipment volumes vary dramatically across HS categories. Rather than synthetic oversampling methods such as SMOTE, the authors adopt stratified up‑sampling, duplicating minority‑class instances proportionally to their mean and median frequencies. This approach preserves the original data distribution, avoids over‑fitting to synthetic samples, and only modestly increases the total record count (from 815,264 to 818,048).

Three deep‑learning architectures are evaluated: Text‑CNN, LSTM, and a fully connected DNN. Hyper‑parameters for each model are tuned via Bayesian optimization, and the Text‑CNN emerges as the best performer, achieving 98 % accuracy on a held‑out test set while maintaining low inference latency. LSTM and DNN models achieve respectable scores but lag behind in both accuracy and response time, making them less suitable for production where deterministic latency is required.

The operational core of the work is a serverless MLOps pipeline built on Amazon Web Services (AWS). Data ingestion and preprocessing are triggered by S3 events and processed by AWS Lambda functions. Model training is orchestrated through AWS Step Functions, which launch SageMaker training jobs with the selected hyper‑parameters. Upon completion, model artifacts are stored in S3, and a Lambda‑based deployment script creates or updates SageMaker endpoints. Real‑time inference is exposed via API Gateway and Lambda, with auto‑scaling policies that adjust compute capacity based on request volume, thereby guaranteeing SLA compliance (e.g., sub‑200 ms latency under peak load).

A distinguishing feature is automated A/B testing. Multiple model versions are simultaneously routed to live traffic using weighted traffic splitting. CloudWatch metrics capture accuracy, latency, and cost per inference; models that meet predefined thresholds (e.g., ≥95 % accuracy, ≤200 ms latency, acceptable cost) are automatically promoted to production, while under‑performing versions are rolled back. Continuous monitoring detects data drift or model performance degradation, triggering a retraining workflow without human intervention.

Cost analysis reveals that transformer‑based models, while capable of matching the Text‑CNN’s accuracy, incur substantially higher training and inference expenses due to GPU requirements and larger memory footprints. Consequently, the authors advocate a “light‑weight‑first” strategy: deploy efficient models like Text‑CNN by default and retain the ability to plug in transformer or LLM variants when specific use‑cases demand higher flexibility or explainability.

The paper concludes that the proposed serverless MLOps architecture delivers reproducibility, auditability, and scalability while minimizing operational overhead. It provides a replicable blueprint for enterprises seeking to industrialize ML in regulated domains. Future work includes integrating Neural Architecture Search (NAS) for automated model design, leveraging large language models for richer explanations, and extending the deployment across multiple AWS regions to further improve latency and fault tolerance.