Flag the review
A fine-tuned DistilBERT model (with a TF–IDF fallback) scores every review against a recall-targeted threshold.
Serious complaints are not missed.
Why HomeLens
Know which providers need attention, and why.
HomeLens AI reads customer reviews of plumbers, electricians, HVAC and other home-service providers, flags the complaints that matter, shows the exact words behind each signal, and tells you what to do next.
Customer review
The technician was three hours late, charged more than the quote, and the repair failed again two days later.
Evidence sentencePricing · sentence risk 0.96
17,943Reviews analyzedTucson, AZ, 2006–2022
254Providers monitoredAnonymized, 20+ reviews each
8Complaint aspectsPricing, safety, reliability and more
0.973Macro-F1, held-out testFine-tuned DistilBERT, businesses held out
Metrics come from the validated course notebook. Reviews are historical customer allegations, not verified findings.
00 — Try it
Paste a review. See the signal.
This is the working analyzer, not a mockup. Pick an example or paste your own text.
Scored in your browser session by the same model the demo uses. Nothing is stored.
Management attentionTF–IDF Logistic Regression
Review risk 0.96Threshold 0.585
PricingTimelinessWorkmanship
The technician was three hours late, charged more than the quote, and the repair failed again two days later.
Recommended action. Review estimate accuracy, change-order approvals, and fee disclosure.
What it does
From a wall of reviews to a short list of what to do.
Three layers turn free-form text into decisions a manager can act on, and every one of them is explainable.
Explain the complaint
A phrase dictionary over eight aspects shows what went wrong.
The evidence sentence is highlighted, so no score is a black box.
Watch the provider
Provider scores, yearly trends and relative concern tiers show who needs management attention over time.
A minimum of 20 reviews each.
01 — Provider monitoring
An interactive risk leaderboard.
Rank 254 anonymized providers by risk score, filter by tier, search by code, and open any row to see its yearly trajectory, signals and recommended operational action.
- Four relative concern tiers, always labeled
- Recent-vs-prior trend with a minimum review count
- Click a row to open triage actions
02 — Explainable by design
See the sentence behind the score.
Every flag comes with the sentence that triggered it, the dictionary phrases that matched, and the terms that moved the model. Managers can disagree with a signal because they can read exactly why it fired.
- Evidence sentence with highlighted phrases
- Term-level weights from the served model
- No dictionary match? It goes to manual review
03 — Complaint aspects
Eight things customers complain about.
Each aspect has its own phrase dictionary and a measured lift: how much more often a review is negative when the aspect is mentioned.
Professionalism
2.8×lift806reviews
very ruderude staffunprofessional
Communication
2.6×lift245reviews
never called backdid not call backdidn't call back
Reliability
2.6×lift378reviews
no shownever showeddid not show
Timeliness
2.4×lift158reviews
arrived lateshowed up latehours late
Warranty
2.0×lift100reviews
warranty not honoredwould not honorrefused to fix
Pricing
2.0×lift276reviews
overchargedcharged morehidden fee
Workmanship
1.9×lift281reviews
poor qualitybad jobterrible job
Safety
1.9×lift136reviews
unsafedangerousfire hazard
Responsible AI
Signals for people, not verdicts.
How we describe outputs
- Review-based risk signal
- Management attention
- Operational concern and complaint signal
- Relative concern tier
- A prompt for human investigation
What we never claim
- Bankruptcy or financial-failure prediction
- Legal liability
- Verified misconduct or fraud
- Verified safety violation
See it on real review data.
The live demo runs on anonymized, historical reviews from Tucson, AZ. No sign-up.