AISummary
Written automatically from the analysis. Click a finding to see the messages behind it.
Top issues
Grouped automatically by meaning. Click one to read what customers wrote.
Volume & mood over time
Messages per day, and how many are complaints
What customers say
Most recent complaints
Compare companies
Share of each company's complaints per issue. Hover a cell for details, click it to read the messages.
1 · Sentiment
A DistilBERT transformer fine-tuned on 11,500 labelled customer messages reads each one as negative, neutral or positive.
| Word counting (baseline) | 72.1% |
| Fine-tuned DistilBERT | 77.7% |
Macro F1 on 1,399 test messages
2 · Issues, found automatically
Each complaint becomes a 384-number "meaning" vector; similar meanings are grouped and named from their most distinctive words. No list of topics is given in advance, so it works on any business's reviews.
| Word-based grouping | NMI 0.18 |
| Meaning-based grouping | NMI 0.25 |
Agreement with human complaint labels
3 · Private by design
Both models run inside your browser. When you upload a file, your customers' messages are analysed on your computer and never sent anywhere.
Limits: English only, sarcasm can fool it, and topic names are suggestions: rename them to your words.