SKIRA: Review Forensics on AMD Token Factory

SKIRA (Review Forensics) is a review-intelligence dashboard for online sellers. Instead of reporting "2.7 stars", it explains the story behind the number. Each review is classified by DeepSeek-V4-Flash on AMD Token Factory, which returns sentiment, confidence, complaint category, the product issue and a recommended action. If the cloud endpoint is unavailable, a rule-based CPU fallback answers and the interface states which engine was used. The dashboard answers what sellers actually ask: what are customers praising and complaining about, which product is getting worse, and what should I check first? Overview shows rating and sentiment breakdowns and topic counts. Reviews is a searchable, filterable list so every number traces back to real text. Alerts compares each topic's share of reviews in the last 7 days with the previous 28 days and flags a significant rise with supporting quotes. Momentum compares each product's last 28 days with the 28 days before. A try-it panel classifies any pasted review. The AMD Benchmark mode runs on 10,000 locally generated synthetic reviews that include sarcasm, negation, rating and text mismatch, and prompt injection edge cases. They are not real customer reviews, their labels are unverified, and platform, product, price and date are simulated. One complaint spike is planted on purpose to test the detector, and the interface labels all of this on screen. A 20-review smoke test (18 via AMD Token Factory, 2 via CPU fallback) checked the pipeline end to end but does not measure real-world accuracy. The backend is Python with SQLite and the standard library HTTP server, so it is easy to deploy. Code: https://github.com/spoorthi-lavu/AMD-ACT3