Enterprise AI Logistics Classifier & Automated Support Email Engine
- 1Independent Researcher
Abstract
Automating customer support tickets in the logistics domain requires high semantic understanding to accurately interpret unstructured user queries, alongside strict backend validation to ensure data security and integrity. This preprint presents the Enterprise-AI-Logistics-Classifier-Automated-Support-Email-Engine, an advanced hybrid pipeline that integrates a fine-tuned RoBERTa (Robustly Optimized BERT Pretraining Approach) deep learning model with deterministic rule-based checks and SQLite database lookups.
The system processes incoming customer messages, performs deep contextual intent classification to categorize queries (such as missing, malformed, invalid, or mismatched parcel numbers), validates tracking information against backend parcel records, and dynamically generates professional, personalized reply templates. Evaluated on a structured test dataset (test.xlsx), the hybrid RoBERTa architecture demonstrated exceptional performance, achieving 100% classification accuracy, precision, recall, and F1-score, with a competitive processing latency (~0.57 seconds per ticket). This repository provides a fully reproducible framework for scalable, AI-driven enterprise customer service operations.
Key Features & Technical Architecture
RoBERTa-Powered Semantic Classification: Utilizes a transformer-based RoBERTa model to capture deep contextual representations of customer support inquiries, moving beyond simple keyword matching to accurately classify complex or ambiguous text inputs.
Hybrid Rule-Based Validation Layer: Combines deep learning predictions with deterministic constraints (e.g., checking empty fields, validating 8-digit parcel constraints) to mitigate model hallucinations and enforce structural logic.
Backend Database Integration: Interacts with a SQLite database (parcel_db.sqlite) to cross-reference parcel numbers against real records, extracting customer identifiers and tracking statuses (In Transit, Delivered, Delayed).
Automated Template Engine: Automatically maps classification outcomes and live database states to professional response templates, ensuring rapid, personalized communication loops.
Open Science Alignment: Built for full transparency and reproducibility, providing training datasets (train.xlsx), evaluation datasets (test.xlsx), modular source code (main.py), and execution logs (results.xlsx).
Methodology & Model Pipeline
Input Ingestion & Tokenization: Raw customer support queries from test.xlsx are tokenized using the RoBERTa tokenizer to preserve sub-word semantics and contextual nuances.
Contextual Inference: The fine-tuned RoBERTa classifier evaluates the text sequence to predict primary intent categories (e.g., valid query vs. missing parcel number).
Database & Rule Validation: The extracted identifiers pass through a multi-tier logic gate:
Missing/Malformed Check: Handled via rule-based filters.
Database Lookup: Queried against parcel_db.sqlite to handle valid matches, invalid numbers, or cross-customer mismatches.
Dynamic Response Assembly: Validated outcomes trigger the corresponding markdown/text reply template, injecting live tracking variables before writing the final audit trail to results.xlsx.
Evaluation Metrics & Results
Metrics Tracked: Accuracy, Precision, Recall, F1-Score, Confusion Matrix, and Inference Latency.
Performance: Achieved a 100% success rate across all core classification metrics on the evaluation dataset.
Efficiency: Maintained an average latency of ~0.57 seconds per ticket (processing a 20-ticket batch in 11.36 seconds, translating to a throughput of ~1.76 tickets/second).
Keywords
RoBERTa; Transformer Models; Logistics Automation; Customer Support; Natural Language Processing; Hybrid AI Pipelines; Open Science
Repository Structure & Software Details
Programming Language: Python
Core Libraries: PyTorch, Transformers (Hugging Face / RoBERTa), Pandas, SQLite3
Repository Link: GitHub - VasudevanS1906/Enterprise-AI-Logistics-Classifier-Automated-Support-Email-Engine
Related
- Autonomous Relationship Engine: Persistent Intent, Goal-State Evaluation, and Cognitive Memory for Enterprise Agentic Execution
- Automated Database Backup and Machine Learning Training Integration via Telegram Bot API in Django Enterprise Systems
- A support vector machine classifier of PM₂.₅-based air quality index categories from co-pollutants and meteorological variables in Ahvaz, Iran
- C-MAGE: Automated Extraction and Organization Engine for Molecular Representations in Chemical Documents
- Early Detection of Mortality Risk in Preterm Neonates Using a Random Forest Classifier and Tree-Based Pipeline Optimization Tool Automated Machine Learning
BibTeX
@misc{swornampillai2026enterprise,
title = {{Enterprise AI Logistics Classifier & Automated Support Email Engine}},
author = {Vasudevan Swornampillai},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23465855},
url = {https://zenodo.org/records/23465855},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Enterprise AI Logistics Classifier & Automated Support Email Engine
AU - Vasudevan Swornampillai
PY - 2026
DA - 2026/10/19
PB - Zenodo
DO - 10.5281/zenodo.23465855
UR - https://zenodo.org/records/23465855
AB - Automating customer support tickets in the logistics domain requires high semantic understanding to accurately interpret unstructured user queries, alongside strict backend validation to ensure data security and integrity. This preprint presents the Enterprise-AI-Logistics-Classifier-Automated-Support-Email-Engine, an advanced hybrid pipeline that integrates a fine-tuned RoBERTa (Robustly Optimized BERT Pretraining Approach) deep learning model with deterministic rule-based checks and SQLite database lookups. The system processes incoming customer messages, performs deep contextual intent classification to categorize queries (such as missing, malformed, invalid, or mismatched parcel numbers), validates tracking information against backend parcel records, and dynamically generates professional, personalized reply templates. Evaluated on a structured test dataset (test.xlsx), the hybrid RoBERTa architecture demonstrated exceptional performance, achieving 100% classification accuracy, precision, recall, and F1-score, with a competitive processing latency (~0.57 seconds per ticket). This repository provides a fully reproducible framework for scalable, AI-driven enterprise customer service operations. Key Features & Technical Architecture RoBERTa-Powered Semantic Classification: Utilizes a transformer-based RoBERTa model to capture deep contextual representations of customer support inquiries, moving beyond simple keyword matching to accurately classify complex or ambiguous text inputs. Hybrid Rule-Based Validation Layer: Combines deep learning predictions with deterministic constraints (e.g., checking empty fields, validating 8-digit parcel constraints) to mitigate model hallucinations and enforce structural logic. Backend Database Integration: Interacts with a SQLite database (parcel_db.sqlite) to cross-reference parcel numbers against real records, extracting customer identifiers and tracking statuses (In Transit, Delivered, Delayed). Automated Template Engine: Automatically maps classification outcomes and live database states to professional response templates, ensuring rapid, personalized communication loops. Open Science Alignment: Built for full transparency and reproducibility, providing training datasets (train.xlsx), evaluation datasets (test.xlsx), modular source code (main.py), and execution logs (results.xlsx). Methodology & Model Pipeline Input Ingestion & Tokenization: Raw customer support queries from test.xlsx are tokenized using the RoBERTa tokenizer to preserve sub-word semantics and contextual nuances. Contextual Inference: The fine-tuned RoBERTa classifier evaluates the text sequence to predict primary intent categories (e.g., valid query vs. missing parcel number). Database & Rule Validation: The extracted identifiers pass through a multi-tier logic gate: Missing/Malformed Check: Handled via rule-based filters. Database Lookup: Queried against parcel_db.sqlite to handle valid matches, invalid numbers, or cross-customer mismatches. Dynamic Response Assembly: Validated outcomes trigger the corresponding markdown/text reply template, injecting live tracking variables before writing the final audit trail to results.xlsx. Evaluation Metrics & Results Metrics Tracked: Accuracy, Precision, Recall, F1-Score, Confusion Matrix, and Inference Latency. Performance: Achieved a 100% success rate across all core classification metrics on the evaluation dataset. Efficiency: Maintained an average latency of ~0.57 seconds per ticket (processing a 20-ticket batch in 11.36 seconds, translating to a throughput of ~1.76 tickets/second). Keywords RoBERTa; Transformer Models; Logistics Automation; Customer Support; Natural Language Processing; Hybrid AI Pipelines; Open Science Repository Structure & Software Details Programming Language: Python Core Libraries: PyTorch, Transformers (Hugging Face / RoBERTa), Pandas, SQLite3 Repository Link: GitHub - VasudevanS1906/Enterprise-AI-Logistics-Classifier-Automated-Support-Email-Engine
ER -