How to build multi-location review classification at scale

The 4-skill bundle on the review-classification agent

One agent. Four coordinated skills. The Ingest + Classify + Calibrate + Audit bundle runs above the review- management + per-platform-review + LLM-classifier + NLP + Bayesian + warehouse substrate and writes one canonical per-location per-review per-label classification record.

Compliance overlay

Five anchors run per-review per-label before any classification distributes to downstream decision systems. The first anchor is operationally distinctive: multi-label classification + per-label confidence calibration + per- label drift detection + per-vertical taxonomy + per- vertical incident reporting (Joint Commission Sentinel Event + FDA MedWatch + DEA + OSHA + TTB + state- board) converge on every per-review classification decision.

FAQ

What is multi-location review classification at scale — and what is the multi-label-classification-times-per-label-confidence-calibration-times-per-label-drift-times-per-vertical-taxonomy-times-Joint-Commission-Sentinel-Event-times-FDA-MedWatch problem distinctive to this skill?

A multi-location operator with 50-300 stores ships per-location per-review classification across 10-15 review platforms (Google Business Profile + Yelp + Apple Maps + Facebook + TripAdvisor + Booking.com + Better Business Bureau + Trustpilot + Glassdoor + Indeed). Per-week per-portfolio: 5,000-50,000 reviews requiring multi-label classification across sentiment + aspect + topic + intent + complaint-class + fraud + per-vertical regulatory-trigger. The four-skill bundle on the review-classification agent — Ingest, Classify, Calibrate, Audit — sits above the review-management + per-platform-review + LLM-classifier + NLP + Bayesian + warehouse substrate (Birdeye + Reputation.com + Podium + GatherUp + ReviewTrackers + OpenAI + Anthropic + Google Gemini + Mistral + Cohere + Meta Llama + Hugging Face Transformers + spaCy + NLTK + scikit-learn + XGBoost + PyMC + Stan) and writes a per-location per-review per-label canonical classification record. The operationally distinctive anchor: per-location per-review multi-label classification (per-review sentiment NPS-aligned + Plutchik emotion wheel + Russell circumplex valence-arousal + per-review aspect (service + product + price + staff + cleanliness + speed + accuracy + accessibility + wait-time + atmosphere + parking + location + value) + per-review topic (per-vertical topic taxonomy + LDA + BERTopic + Top2Vec + dynamic-topic-model) + per-review intent (complaint + compliment + question + suggestion + recall + warranty-claim + retention-signal + churn-signal) + per-review complaint-class (per-vertical regulatory-trigger + per-vertical legal-risk + per-vertical safety-incident + per-vertical professional-licensing-board reportable) + per-review fraud-detection (per-review fake-review signal via FTC Fake Review Rule 16 CFR Part 465 + per-Google fake-review policy + per-Yelp fake-review policy + per-reviewer-history) + per-vertical NPS-impact + per-vertical retention-impact + per-vertical CSAT impact + per-vertical CES Customer Effort Score impact) + zero-shot + few-shot + LLM-as-judge + RLHF + DPO + per-label confidence calibration (Brier + ECE + MCE + reliability diagram + isotonic + Platt + temperature scaling + conformal + Mondrian conformal + ACI) + per-label drift detection (PSI + KL divergence + Jensen-Shannon + Wasserstein + MMD) + per-vertical taxonomy (ICD-10 + SNOMED-CT + LOINC + RxNorm healthcare + NAICS + SIC + GICS + UNSPSC) + per-vertical incident reporting (Joint Commission Sentinel Event Database + FDA MedWatch adverse event MAUDE + DEA per-state pharmacy + alcohol TTB enforcement + state-board enforcement + OSHA incident + EPA toxic-release + per-state consumer protection).

Why do Birdeye + Reputation.com + Podium + Brandwatch + OpenAI + Hugging Face Transformers break at multi-location-multi-platform-multi-label-classification scale?

Each review-management + LLM-classifier + NLP vendor ships per-account flat single-label classification primitive at single-platform level. None coordinates per-location per-review multi-label classification (sentiment + aspect + topic + intent + complaint-class + fraud + per-vertical regulatory-trigger) + zero-shot + few-shot + LLM-as-judge + RLHF + DPO + per-label confidence calibration (Brier + ECE + isotonic + Platt + conformal) + per-label drift detection (PSI + KL + JS + Wasserstein + MMD) + per-vertical taxonomy + per-vertical incident reporting simultaneously. None handles Joint Commission Sentinel Event + FDA MedWatch adverse event MAUDE + DEA per-state pharmacy + alcohol TTB enforcement + state-board enforcement + OSHA incident reporting convergence. None gates against FTC Fake Review Rule + ABA Model Rule + HIPAA + FINRA + Tennessee ELVIS Act + per-state biometric. None enforces SOX 302/404/906 when public-company franchisor material. None writes a per-location per-review per-label WORM classification audit trail. The four-skill bundle Ingest + Classify + Calibrate + Audit sits above the review-management + per-platform-review + LLM-classifier + NLP + Bayesian + warehouse substrate — it does not replace it.

How does Ingest + Classify work?

Ingest runs per-platform per-location per-review ingestion across Google Business Profile + Yelp Fusion + Apple Maps Business Connect + Facebook + TripAdvisor + Booking.com + Better Business Bureau + Trustpilot + Glassdoor + Indeed with per-vendor DPA + per-API rate-limit + per-source ToS compliance. Per-review normalization + per-review per-platform metadata (per-reviewer-history + verified-purchase tag + review-photo + review-video). Classify runs per-location per-review multi-label classification: per-review sentiment (NPS-aligned -100 to +100 + Plutchik emotion wheel + Russell circumplex) + per-review aspect (service + product + price + staff + cleanliness + speed + accuracy + accessibility + wait-time + atmosphere + parking + location + value) via spaCy + NLTK + Stanza + Flair + AllenNLP + Hugging Face Transformers aspect-based sentiment analysis (ABSA). Per-review topic classification via LDA + BERTopic + Top2Vec + dynamic-topic-model. Per-review intent classification (complaint + compliment + question + suggestion + recall + warranty-claim + retention-signal + churn-signal) via LLM zero-shot + few-shot + LLM-as-judge + RLHF + DPO Direct Preference Optimization. Per-review complaint-class (per-vertical regulatory-trigger + per-vertical legal-risk + per-vertical safety-incident + per-vertical professional-licensing-board reportable). Per-review fraud-detection via FTC Fake Review Rule 16 CFR Part 465 signal + per-Google fake-review policy + per-Yelp fake-review policy + per-reviewer-history. Per-vertical taxonomy alignment (ICD-10 + SNOMED-CT + LOINC + RxNorm + NAICS + SIC + GICS + UNSPSC).

What does Calibrate + Audit do?

Calibrate runs per-label confidence calibration: per-label Brier score + Expected Calibration Error (ECE) + Maximum Calibration Error (MCE) + reliability diagram + isotonic regression + Platt scaling + temperature scaling + conformal prediction intervals + Mondrian conformal (per-label bin) + Adaptive conformal inference (ACI). Per-label drift detection: per-label Population Stability Index (PSI) + KL divergence + Jensen-Shannon divergence + Wasserstein distance + Maximum Mean Discrepancy (MMD). Per-vertical replication-crisis statistical discipline: per-classifier Brier-decomposition (reliability + resolution + uncertainty) + per-classifier proper scoring rules (logarithmic + Brier + spherical + RPS + CRPS + quantile + pinball) + per-classifier Rosenbaum sensitivity (Γ) + E-value (VanderWeele Ding 2017) + Bonferroni + Benjamini-Hochberg FDR + Holm-Bonferroni + Šidák multi-comparison correction. Per-review per-label severity classification: P0 per-vertical regulatory-trigger (Joint Commission Sentinel Event + FDA MedWatch adverse-event + DEA pharmacy violation + OSHA incident + alcohol TTB enforcement + state-board enforcement) (immediate alert + escalation) + P1 ABA Model Rule violation + per-state attorney advertising 72-hour + P2 fraud-class fail (FTC Fake Review Rule violation) 7-day + P3 per-label calibration drift 30-day + P4 docs-only. Gate runs 5 anchors per-review per-label before any classification distributes to downstream decision systems. (1) Multi-label classification + per-label confidence calibration + per-label drift detection + per-vertical taxonomy + per-vertical incident reporting (Joint Commission + MedWatch + DEA + OSHA + TTB + state-board) + replication-crisis discipline. (2) FTC Fake Review Rule 16 CFR Part 465 + FTC Endorsement Guides + Section 5 + Pfizer 1972 + CFPB UDAAP + Lanham + USPTO + Robinson-Patman + FDD Item 12. (3) ABA Model Rule 7.1-7.5 + 1.18 + per-state attorney advertising 50-state + HIPAA 45 CFR 164.502/504/514 Safe Harbor when health/wellness review + state medical board + FINRA Rule 2210 + SEC Regulation FD + per-state professional licensing + FDA OPDP + DEA + alcohol + + tobacco. (4) Tennessee ELVIS Act 2024 right-of-publicity when AI-classified voice/likeness + per-state right-of-publicity + EU AI Act Article 50 transparency when AI-ML review classification + Article 13/14/15 + Annex III when AI-ML review classification drives escalation routing + Article 6/27 FRIA + DSA + DMA + GDPR Article 6/7/22/28/30 + LGPD + DPDP + PIPEDA + Quebec Law 25 + CCPA + CPRA + 18-state + per-state biometric (BIPA + CUBI + Washington biometric). (5) WCAG 2.2 AA + ARIA + EAA + ADA Title III + Section 508 + Section 230 + DMCA Section 512 + SOX 302/404/906 + COSO + Exchange Act 13(b)(2) + SEC Reg S-K. Audit writes a per-location per-review per-label WORM classification record: per-review snapshot + per-label classification + per-label confidence calibration + per-label drift + per-vertical taxonomy alignment + per-vertical incident-report flag + per-anchor gate-pass + AI-ML provenance + EU AI Act FRIA. Retention: 7-year FTC + 7-year IRS + 7-year HIPAA + 7-year state bar + 6-year SEC + 3-year FINRA + 7-year SOX + GDPR Article 30 + EU AI Act Article 12 + SOC 2 CC7/CC8.

What does this skill connect to on the review-classification agent and across the swarm?

On the review-classification agent: per-location per-review multi-label classification + per-label confidence calibration + per-label drift detection + per-vertical incident reporting. Across the swarm: auto-publish gating for AI review-responses (#606 DOWNSTREAM consumer of canonical per-review classification) + per-location AI review-response drafting (#565 DOWNSTREAM consumer of per-review classification) + per-location AI-calibrated forecasting (#600 same per-label calibration substrate) + per-location per-cohort two-sigma anomaly detection (#608 same multi-comparison correction + replication-crisis discipline substrate) + per-vertical compliance overlay (#615 DOWNSTREAM consumer of per-vertical incident-reporting + per-vertical regulatory-trigger) + integration-drift-monitor agent (#562 + #569 + #570) + per-state-overlay-composer (#599 UPSTREAM canonical for ABA + per-state attorney advertising + HIPAA + FINRA per-state). Commercial-pillar parent: /review-response-orchestration.

What does the 6-workstream pre-engagement-baseline reporting cycle look like for this skill?

Every two weeks during the Tier 3 Fractional CMO with AI Swarm engagement, six workstreams report against the pre-engagement baseline. Workstream 1: per-portfolio per-location per-review per-label review-classification coverage — locations covered + platforms ingested + reviews classified. Workstream 2: Ingest per-platform per-location per-review ingestion flow — per-vendor DPA + rate-limit + ToS compliance + per-review per-platform metadata + per-reviewer-history. Workstream 3: Classify per-location per-review multi-label classification flow — sentiment + aspect + topic + intent + complaint-class + fraud-detection + per-vertical regulatory-trigger + per-vertical taxonomy alignment. Workstream 4: Calibrate per-label confidence calibration + per-label drift detection flow — Brier + ECE + isotonic + Platt + conformal + PSI + KL + JS + Wasserstein + MMD + replication-crisis discipline. Workstream 5: Regulatory-defense audit coverage — multi-label classification + calibration + drift + per-vertical taxonomy + per-vertical incident reporting (Joint Commission + MedWatch + DEA + OSHA) + FTC Fake Review + HIPAA + ABA + ELVIS Act + EU AI Act Article 50 + SOX. Workstream 6: FBC feedback-loop pattern-learning — per-review per-label realized-vs-predicted classification + per-vertical regulatory-trigger enforcement retrospective + per-label drift retrospective.

Related reading

• Parent commercial pillar: review response orchestration
• Sibling build-pillar: auto-publish gating for AI review-responses (#606 DOWNSTREAM consumer of canonical per-review classification)
• Sibling build-pillar: per-location AI review-response drafting (#565 DOWNSTREAM consumer of per-review classification)
• Sibling build-pillar: per-vertical compliance overlay (#615 DOWNSTREAM consumer of per-vertical incident- reporting + per-vertical regulatory-trigger)
• Fractional CMO with AI Swarm
• AI Readiness Assessment