Software Development
HazardWatch
A product safety early warning system that reads customer reviews and flags the ones describing real hazards. A fine tuned BERT + BiGRU classifier scores each review, a MiniLM and FAISS retrieval index surfaces similar past incidents, and product level aggregation ranks the riskiest listings. Served as a FastAPI API with a browser prototype for live feed monitoring.
Problem
- Sellers and regulators can’t easily see which products are most dangerous.
- Manual checks can’t keep up with daily reviews.
- Keyword filters miss hidden risks and flag harmless reviews.
- Repeat complaints about the same product go unnoticed.
Solution
- BERT + BiGRU scores each review for safety risks.
- Using FAISS, similar flagged reviews are matched to find repeated risks.
- Duplicate reviews are filtered to avoid repeated reports.
- Secure API access handles new and bulk review submissions.
- Products are ranked based on the number and severity of reports.
- A browser dashboard shows live alerts and product risk levels.
Impact
- Reviewers see a shortlist of likely safety risks.
- Similar cases show whether a problem is recurring.
- Risk rankings highlight products needing the most attention.
- Live feeds show how the system handles reviews in real time.
- Runs without a GPU, keeping deployment costs low.
