Public safety & social data
Production classification API
A fine-tuned language model that classifies short social text for emergency relevance—benchmarked rigorously, ranked competitively on a public leaderboard, and deployed as a real-time API.
- Fine-tuning
- API
- MLOps
- NLP
Context
Emergency services need fast, reliable signals in noisy public text streams—but off-the-shelf models rarely match the specificity required. The goal was to train a classifier that could generalize well on real-world short text, compare architectural options honestly, and ship the winner as a production endpoint others could call.
What we built
We fine-tuned a BERT transformer on labeled short-text data, prototyped LSTM and RNN alternatives for comparison, and documented performance trade-offs before selecting the production architecture. The final model was packaged and deployed on managed cloud ML infrastructure as a low-latency prediction API.
Delivery highlights
- Fine-tuned BERT transformer for binary emergency-relevance classification
- Benchmarked against LSTM and RNN prototypes with documented metrics
- Placed in the top 60 of 700 teams on a public modeling competition
- Deployed the production model as a real-time API on SageMaker
Impact
- Demonstrated strong generalization on held-out and competition evaluation data
- Provided a reproducible training and deployment playbook for similar NLP tasks
- Delivered an inference endpoint suitable for integration into downstream systems