What to review
Prototype LLM Insights report for Thunderbird Android support questions — June 2026 vs May 2026. This is the AI counterpart to Project 1. I ran the no-AI (Project 1) approach first and the LLM approach second so you can compare them on the same data.
Rendered reports
Underlying data (CSVs on GitHub)
Per-question LLM labels — theme, category, severity, root-cause hypothesis, verbatim quote, and value signals (solved / trusted-last-answer / answered):
Non-AI feature tables (regex tags used by Project 1):
Key finding
At ~46 questions/month, the non-AI detector flags nothing (0 spikes / 0 version×cause — its thresholds are tuned for desktop's ~700/mo); only low-count mix tables render. The LLM approach surfaces real, actionable signal from the same 46 questions:
- Stopped receiving new mail on Android — mean severity 4.0, 33% unanswered; hypothesis: IMAP IDLE / doze battery-optimization halting background sync.
- No new-mail notifications despite sync — Android notification-pipeline gap.
- Save All attachments → overwritten / zero-byte files — a concrete data-loss bug.
- 5 new mobile/desktop-parity clusters (local/POP3 folders on Android, flagged-mail view, password-change discoverability, …).
Cost: $0.65 total (classify $0.53 + reduce $0.12).
Questions for you
- Are the ranked issues + root-cause hypotheses credible and useful to engineering?
- Worth automating for Android (daily/monthly, like the desktop plan)?
- Any format changes before sharing more widely?
Prototype — an LLM-derived triage signal over free-text support questions, not proof. Counts are exact (computed in Python); clustering and prose are LLM-derived.
What to review
Prototype LLM Insights report for Thunderbird Android support questions — June 2026 vs May 2026. This is the AI counterpart to Project 1. I ran the no-AI (Project 1) approach first and the LLM approach second so you can compare them on the same data.
Rendered reports
Underlying data (CSVs on GitHub)
Per-question LLM labels — theme, category, severity, root-cause hypothesis, verbatim quote, and value signals (solved / trusted-last-answer / answered):
Non-AI feature tables (regex tags used by Project 1):
Key finding
At ~46 questions/month, the non-AI detector flags nothing (0 spikes / 0 version×cause — its thresholds are tuned for desktop's ~700/mo); only low-count mix tables render. The LLM approach surfaces real, actionable signal from the same 46 questions:
Cost: $0.65 total (classify $0.53 + reduce $0.12).
Questions for you
Prototype — an LLM-derived triage signal over free-text support questions, not proof. Counts are exact (computed in Python); clustering and prose are LLM-derived.