Automated Review Rating System
Fine-tuned BERT classifying and prioritizing urban-development feedback — published in SASBE.
- role
- AI Engineer & Team Lead
- duration
- 4 months
- team
- 2
- stack
- Python · BERT · TensorFlow · Plotly
“NLP when NLP was still a thing you built yourself.”
The problem
City planners receive thousands of free-text reviews on urban projects — roads, parks, transit. Reading and triaging that manually means the loudest voices win, not the most important issues.
The approach
Fine-tune a BERT model for the domain — multilingual urban feedback, balanced across positive/negative/neutral — then compose it with a priority-scoring pass (urgency + scope + sentiment) that ranks issues for dispatch. Output feeds a Plotly dashboard with cluster views, trend lines, and drill-downs so planners can argue about specific items, not anecdotes.
A parallel processing pipeline handled the large review corpora without overnight runs.
The stack
TensorFlow + a fine-tuned BERT checkpoint · Python for pipelines and augmentation to rebalance the dataset · Plotly for interactive visualization · TensorBoard during training for loss/metric surfaces.
Reflection
The research lesson was that the model was half the work. The other half was building an interface where a non-ML planner could push back on its verdict — because that’s what makes the predictions useful, not just accurate.