Business Context

AMD Construction Group collects customer feedback from post-project surveys, warranty emails, and AMD BuildSmart support tickets. The customer experience team wants a sentiment analysis system that converts this unstructured text into actionable signal for project quality monitoring and escalation.

Data

You are given roughly 180,000 feedback records from the last 24 months across residential and commercial projects. Text is primarily English (96%), with some Spanish comments. Feedback ranges from 5 to 400 words (median: 42 words) and includes free-form comments, contractor references, schedule complaints, punch-list issues, and praise for crews. Labels are available for a subset of 45,000 records with three classes: negative (28%), neutral (37%), and positive (35%). Around 12% of comments contain boilerplate survey text or duplicated signatures.

Success Criteria

A good solution should achieve macro-F1 ≥ 0.82 on the labeled set and negative-class recall ≥ 0.90, since missed negative feedback can hide quality or service issues. The output should also support downstream aggregation by project, region, and product line.

Constraints

• Inference should run in <150 ms per comment for batch scoring in AMD Construction Group analytics workflows.
• The solution must be explainable enough for operations managers to trust why a comment was flagged negative.
• PII such as names, phone numbers, and addresses should not be retained in model features.

Requirements

1. Build a sentiment analysis pipeline for AMD BuildSmart customer feedback.
2. Describe preprocessing for noisy construction-domain text.
3. Implement a baseline and a transformer-based model in Python.
4. Evaluate class-wise performance, especially negative recall.
5. Explain how you would turn sentiment outputs into business signal such as project-level risk flags or recurring issue themes.