Architecture swarm · Customer-data-graph agent · LTV-math- primitives skill · Build pillar · Published July 11, 2026

How to build per-location per-channel per-brand CLV at runtime

Direct-to-consumer ecommerce and multi-location franchise systems often run several banners across several channels (DTC ecom + Amazon + Walmart Connect + retail + B2B wholesale + subscription) and need a customer-lifetime-value score that respects the per-location, per-channel, and per-brand grain in real time. Klaviyo Predictive CLV, Lifetimes Python, Theta Equity, Retention Science, Optimove, Custora, Glew, Polar Analytics, Tresl, Daasity, Twilio Engage, Segment, Amplitude, and Hightouch ship per-account-level CLV predictions that work well when the customer touches one brand on one channel; they break the moment a single customer transacts across multiple grains. The Cohort + Recalibrate + Gate + Audit skill bundle on the customer-data-graph agent sits above the vendor layer and exposes a runtime per-location per-channel per-brand per- customer CLV with confidence interval, explainability trace, and named regulatory anchors preserved in every audit record: CCPA right to opt out of automated decision-making + EU AI Act Article 22 + Annex III + FCRA + ECOA / Fair Housing disparate- impact + replication-crisis calibration discipline + FASB ASC 606 + SEC Regulation G non-GAAP + NIST AI RMF.

Start with a Tier 1 AI Readiness Assessment Tier 3 Fractional CMO with AI Swarm Take the 3-question quiz

The 4-skill bundle on the customer-data-graph agent

Cohort

Per-customer cohort assignment across per-location, per- channel, per-brand, per-vertical, per-tenure (0–30 day + 30–90 + 90–180 + 180–365 + 365-day-plus), per-RFM, per-seasonality, per-loyalty-tier, per-subscription- status, per-acquisition-source, per-first-order-SKU. Per-cohort CLV model ensemble: probabilistic (BG/NBD + Pareto/NBD + Gamma-Gamma + MBG/NBD + CDNOW Schmittlein- Morrison-Colombo via Lifetimes Python + PyMC-Marketing + CLVTools-R + BTYDplus-R + Stan + brms-R) + deep-learning (LSTM + Transformer + Convolutional Sequence-to-Sequence in scikit-learn + PyTorch + JAX + TensorFlow + Hugging Face Transformers) with stacking + blending + cross-validation + hyperparameter tuning + A/B testing + MLflow + Weights & Biases + CometML + Neptune.ai + ClearML experiment tracking. Per-customer embedding stored in Pinecone + Weaviate + Qdrant + Milvus + Chroma + pgvector + Vertex AI Vector Search.

Recalibrate

Per-cohort calibration pipeline: Platt scaling + isotonic regression + temperature scaling + Bayesian binning + expected calibration error (ECE) + reliability diagram + Brier score. Per-cohort daily / weekly / monthly recalibration cadence with FBC pattern learning. Per-customer RFM (recency + frequency + monetary + quintile + percentile + cross-cohort percentile). Confidence tier (high / medium / low / no- confidence) + confidence interval (Bayesian posterior + frequentist + bootstrap). Runtime lookup served from Redis + DynamoDB + Cassandra + ScyllaDB + Aerospike cache with TTL + cache-warm + invalidation; real-time + batch + streaming recompute (Apache Flink + Apache Spark Streaming + Materialize + RisingWave); runtime latency p50 / p95 / p99.

Gate

Five anchors before any runtime CLV writes downstream. CCPA + CPRA right to opt out of automated decision-making + 17- state + WA My Health My Data + Texas SCOPE + GDPR Article 22 + 6 + 7 + 9 + 17 + LGPD + DPDP + PIPEDA + Quebec Law 25 + EU AI Act Article 22 + 26 + 50 + 13 + 14 + 15 + Annex III high- risk classification. FCRA + ECOA Regulation B disparate- impact + Fair Housing Act + GLBA + COPPA. Replication-crisis calibration discipline (Brier score + ECE + reliability diagram + Platt + isotonic + temperature + Bayesian binning + multiple-comparisons correction). FASB ASC 606 + FASB ASC 842 + SEC Regulation G non-GAAP + SEC C&DI Q100 / 101 / 102 + AICPA non-GAAP + PCAOB AS 2410. FTC Section 5 + Pfizer 1972 + FTC Endorsement Guides 16 CFR Part 255 + SOC 2 Type II + ISO 27001 + ISO 27701 + NIST AI RMF + ISO 42001 + per- vendor LLM zero-retention. Policy-as-code via OPA Rego + AWS Cedar + Casbin + Cerbos + Oso + Styra DAS + Permit.io.

Audit

Per-customer-CLV WORM record at every Cohort assignment and every Recalibrate run. DSAR export shipped on demand for CCPA Right to Know + CPRA right to correct + GDPR Article 15 + Article 17 + 17-state DSAR. Storage: AWS S3 Object Lock + Azure Blob immutable + Google Cloud Storage Bucket Lock + Wasabi WORM. Retention stacks (longest applicable wins): 7- year FTC + 7-year IRS + 7-year FDD + per-state franchise + GDPR Article 30 + CCPA 12-month look-back + EU AI Act Article 12 + SOX Section 802 7-year + SOC 2 CC7 / CC8. End-to-end replay rewinds every stage with confidence tier and explainability.

The real vendor ecosystem this sits above

CLV providers + numerical stack

Klaviyo Predictive CLV + Lifetimes Python library + PyMC- Marketing + CLVTools-R + BTYDplus-R + Stan + brms-R probabilistic; Theta Equity + Retention Science + Optimove + Custora + Glew + Polar Analytics + Tresl + Daasity vendor CLV; Twilio Engage + Segment Customer Profiles + Amplitude Audiences + Hightouch downstream CDP surface; scikit-learn + PyTorch + JAX + TensorFlow + Hugging Face Transformers deep- learning numerical; MLflow + Weights & Biases + CometML + Neptune.ai + ClearML experiment tracking.

Embedding store + runtime cache + streaming

Embedding stores (Pinecone + Weaviate + Qdrant + Milvus + Chroma + pgvector + Vertex AI Vector Search) hold per- customer purchase-history + clickstream + engagement + SKU embeddings. Runtime cache (Redis + DynamoDB + Cassandra + ScyllaDB + Aerospike) serves the calibrated score. Streaming recompute (Apache Flink + Apache Spark Streaming + Materialize + RisingWave) keeps the score fresh as events flow through. OpenAI + Anthropic LLM tie-breakers under per-vendor zero- retention back explainability narrative; LangSmith + Weights & Biases + Arize + WhyLabs + Helicone + Langfuse + Galileo observability.

Policy-as-code + WORM + sibling skills

OPA Rego + AWS Cedar + Casbin + Cerbos + Oso + Styra DAS + Permit.io policy-as-code expresses every Gate rule. AWS S3 Object Lock + Azure Blob immutable + Google Cloud Storage Bucket Lock + Wasabi compliance WORM holds the per-customer audit substrate. Sibling skills on the customer-data-graph agent: versioned customer history (DSAR substrate); cross- touchpoint identity resolution; deterministic + probabilistic hybrid identity resolution; behavioral-signal ingestion + customer graph; per-field conflict resolution policy.

The 6-workstream reporting cycle

Numeric uplift commitments are not made up-front. The engagement ships a pre-engagement baseline across six workstreams; the cycle tracks delta against that baseline. Reporting is the substrate, not the promise.

Cohort coverage.Per-cohort-axis coverage across the 11 standing cohort axes; model-ensemble coverage across probabilistic + deep-learning families; per-customer embedding completeness; MLflow / Weights & Biases run completeness.
Recalibrate quality. Per-cohort Brier score + expected calibration error + reliability diagram conformance; Platt + isotonic + temperature + Bayesian binning adherence; per-cohort recalibration cadence adherence; runtime cache hit rate; runtime latency p50 / p95 / p99.
Gate quality. Per-anchor evaluation completeness (CCPA opt-out + GDPR Article 22 + EU AI Act Article 22 + Annex III + FCRA + ECOA + Fair Housing + replication-crisis calibration discipline + FASB ASC 606 + SEC Reg G); per-anchor pass / fail / route-to-counsel distribution; per-DSAR fulfillment turnaround; ECOA disparate-impact audit cadence.
Audit quality. Per-customer-CLV WORM record completeness; retention-window coverage (longest of 7-year FTC + 7-year IRS + 7-year FDD + per-state franchise + GDPR Article 30 + CCPA 12-month + EU AI Act Article 12 + SOX 802 + SOC 2 CC7 / CC8); end-to-end replay success rate.
Compliance posture. CCPA right-to-opt-out propagation completeness across downstream consumers; EU AI Act Article 50 disclosure coverage; ECOA disparate-impact audit posture; FASB ASC 606 reconciliation completeness when ARR / cohort-CLV referenced in external reporting; SEC Regulation G reconciliation when public-co or PE-sponsor disclosure.
Audit-trail completeness. Per-anchor regulatory citation completeness; sibling-handoff pointer completeness into the customer-data-graph bundle (versioned customer history + cross-touchpoint identity resolution + deterministic + probabilistic hybrid identity resolution + behavioral-signal ingestion + per-field conflict resolution policy) and into the consuming skills (per-location list-segmentation per-LTV- quintile criterion; monthly-executive-summary-drafting on the rollup-reporting agent for customer / loyalty variance commentary; churn-prediction; tier-transition timing; save- flow propensity; cross-location offer coordination).

Frequently asked questions

What is per-location per-channel per-brand CLV at runtime — and why is account-level Klaviyo Predictive CLV not enough?

Direct-to-consumer ecommerce + multi-location franchise systems often run several banners across several channels (DTC ecom + Amazon + Walmart Connect + retail + B2B wholesale + subscription) and need a customer-lifetime-value score that respects the per-location, per-channel, and per-brand grain in real time. Klaviyo Predictive CLV, the Lifetimes Python library, Theta Equity, Retention Science, Optimove, Custora (now part of Optimove), Glew, Polar Analytics, Tresl, Daasity, Twilio Engage, Segment Customer Profiles, Amplitude Audiences, and Hightouch ship per-account-level CLV predictions that work well when the customer touches one brand on one channel through one storefront. They break the moment a single customer transacts across multiple per-location per-channel per-brand grains. The four-skill bundle on the customer-data-graph agent — Cohort, Recalibrate, Gate, Audit — sits above the vendor CLV layer and exposes a runtime per-location per-channel per-brand per-customer CLV score with confidence interval, explainability trace, and named regulatory citations preserved in the audit trail.

Why do Klaviyo Predictive CLV + Lifetimes + Theta Equity + Optimove + Glew + Polar Analytics + Twilio Engage + Segment + Amplitude + Hightouch break at multi-location multi-channel multi-brand scale?

Each vendor ships a per-account-level CLV primitive. None coordinates cohort assignment across the operator’s per-location, per-channel, per-brand grain (a single customer ends up with multiple disconnected CLV scores across the per-account silos). None runs a per-cohort ensemble of probabilistic CLV models (BG/NBD + Pareto/NBD + Gamma-Gamma + MBG/NBD + CDNOW Schmittlein-Morrison-Colombo + Lifetimes Python + PyMC-Marketing + CLVTools-R + BTYDplus-R) alongside deep-learning (LSTM + Transformer + Convolutional Sequence-to-Sequence). None re-calibrates per cohort on a Brier score + expected calibration error + reliability diagram pipeline. None enforces CCPA + CPRA right to opt out of automated decision-making + GDPR Article 22 + EU AI Act Article 22 transparency of automated decisions before the runtime CLV writes downstream to offer-eligibility, price-personalization, tier-assignment, save-flow routing, churn-prevention. None reconciles ARR + cohort-CLV references against FASB ASC 606 revenue-recognition timing and SEC Regulation G non-GAAP reconciliation when the scores reach investors. The four-skill bundle Cohort + Recalibrate + Gate + Audit sits above the vendor layer — it does not replace it. Cohort assigns each customer to per-location, per-channel, per-brand, per-vertical, per-tenure, per-RFM, per-loyalty-tier, per-subscription-status, per-acquisition-source, per-first-order-SKU cohorts and runs the model ensemble. Recalibrate runs the Platt + isotonic + temperature + Bayesian binning calibration discipline per cohort. Gate enforces the regulatory anchors. Audit writes a per-customer-CLV WORM record.

What does Cohort do — per-customer cohort assignment + per-cohort CLV model ensemble?

Cohort assigns each customer to the standing cohort axes — per-location, per-channel, per-brand, per-vertical, per-tenure (0–30 day + 30–90 + 90–180 + 180–365 + 365-day-plus), per-RFM, per-seasonality, per-loyalty-tier, per-subscription-status, per-acquisition-source, per-first-order-SKU — with cohort-membership versioning that hands off to the behavioral-cohort-computation sibling skill. Per-cohort CLV model ensemble runs probabilistic models (BG/NBD Beta-Geometric/NBD + Pareto/NBD + Gamma-Gamma monetary-value + MBG/NBD Modified Beta-Geometric/NBD + CDNOW Schmittlein-Morrison-Colombo) using Lifetimes Python + PyMC-Marketing + CLVTools-R + BTYDplus-R + Stan + brms-R, alongside deep-learning (LSTM + Transformer + Convolutional Sequence-to-Sequence implemented in scikit-learn + PyTorch + JAX + TensorFlow + Hugging Face Transformers), with model stacking and blending, cross-validation, hyperparameter tuning, A/B testing, and MLflow + Weights & Biases + CometML + Neptune.ai + ClearML experiment tracking. Per-customer embedding (purchase-history + clickstream + engagement + SKU embeddings) stored in Pinecone + Weaviate + Qdrant + Milvus + Chroma + pgvector + Vertex AI Vector Search with drift monitoring. Per-customer confidence tier and explainability trace written into Audit at every score.

What does Recalibrate do — Platt + isotonic + temperature + Bayesian binning + Brier score discipline?

Recalibrate runs the standing calibration pipeline per cohort: Platt scaling for sigmoid recalibration; isotonic regression for monotone non-parametric recalibration; temperature scaling for over-confident neural outputs; Bayesian binning for Bayesian-credible recalibration; expected calibration error (ECE) and adaptive calibration error tracking; reliability diagram for visual diagnosis; Brier score for proper scoring. Recalibration cadence runs per-cohort on a daily / weekly / monthly schedule with FBC pattern learning feeding back into Cohort. Per-customer RFM (recency + frequency + monetary value + RFM quintile + RFM percentile + cross-cohort percentile) feeds the calibration stack. The output is a per-customer per-cohort calibrated CLV with confidence tier (high / medium / low / no-confidence) and a confidence interval (Bayesian posterior + frequentist + bootstrap). Runtime lookup serves the calibrated score from Redis + DynamoDB + Cassandra + ScyllaDB + Aerospike cache with TTL + cache-warm + cache-invalidation policy; real-time recompute, batch recompute, and streaming recompute (Apache Flink + Apache Spark Streaming + Materialize + RisingWave) keep the score fresh; runtime latency tracked at p50 / p95 / p99.

What does Gate do — CCPA opt-out + GDPR Article 22 + EU AI Act Article 22 + Annex III + FCRA + ECOA + Fair Housing + FASB ASC 606 + SEC Reg G non-GAAP?

Gate evaluates five operationally distinctive anchors before any runtime CLV score writes downstream. Anchor 1 (the most operationally distinctive): CCPA + CPRA right to opt out of automated decision-making + 17-state comprehensive privacy (Virginia VCDPA + Colorado CPA + Connecticut CTDPA + Utah UCPA + Texas TDPSA + Florida FDBR + Oregon OCPA + Montana CDPA + Iowa ICDPA + Indiana INCDPA + Tennessee TIPA + Delaware DPDPA + New Hampshire NHPA + New Jersey NJDPA + Maryland MODPA + Minnesota MCDPA + Rhode Island RIDPPA) + Washington My Health My Data Act 2024 + Texas SCOPE Act 2024 + GDPR Article 22 transparency of automated decisions + Article 6 lawful basis + Article 7 consent + Article 9 special categories + Article 17 right to erasure + LGPD + DPDP + PIPEDA + Quebec Law 25; EU AI Act Article 22 + Article 26 deployer obligations + Article 50 transparency for AI-generated content + Article 13 transparency to deployers + Article 14 human oversight + Article 15 accuracy + robustness + cybersecurity + Annex III high-risk classification (when CLV drives offer-eligibility + price-personalization + credit decisioning) — runtime CLV scores ARE automated decisions; opt-out, transparency, and human-oversight requirements apply directly. Anchor 2 (credit + anti-discrimination): FCRA when CLV-driven decisioning affects credit decisioning + prescreen + permissible purpose + adverse-action notice; ECOA Regulation B disparate-impact when CLV proxies for protected class via ZIP code + surname + first-order-SKU + acquisition-source + demographic correlate; Fair Housing Act disparate-impact; GLBA Safeguards Rule when financial-services CLV; COPPA 15 USC 6501 for under-13 cohort. Anchor 3 (replication-crisis calibration discipline): Brier score (proper scoring rule); expected calibration error (ECE); reliability diagram; Platt scaling + isotonic regression + temperature scaling + Bayesian binning for recalibration; per-cohort hold-out testing; pre-registered ablation studies; multiple-comparisons correction (Bonferroni + Benjamini-Hochberg FDR) when reporting per-cohort lift. Anchor 4 (accounting + investor disclosure): FASB ASC 606 revenue recognition 5-step model when CLV references subscription contracts + contract-asset accounting + variable consideration + significant financing component; FASB ASC 842 leases; SEC Regulation G non-GAAP financial measures reconciliation + SEC C&DI Q100 / 101 / 102 when ARR + cohort-CLV reach public-co or PE-sponsor disclosures; AICPA non-GAAP financial measures guidance; PCAOB AS 2410. Anchor 5 (general AI-governance + security): FTC Section 5 + Pfizer 1972 substantiation when CLV scores drive customer-facing claims; FTC Endorsement Guides 16 CFR Part 255 (2023 AI-content); SOC 2 Type II Common Criteria CC2 + CC3 + CC6 + CC7 + CC8; ISO 27001; ISO 27701 privacy information management; NIST AI Risk Management Framework Govern + Map + Measure + Manage; ISO 42001 AI Management System; per-vendor LLM zero-retention verified per call. Policy-as-code expression via OPA Rego + AWS Cedar + Casbin + Cerbos + Oso + Styra DAS + Permit.io.

What does Audit do — per-customer CLV WORM record + DSAR export + end-to-end replay?

Audit writes a per-customer-CLV WORM record at every Cohort assignment and every Recalibrate run: per-customer ID + per-location pointer + per-channel pointer + per-brand pointer + cohort-assignment record (per-location + per-channel + per-brand + per-vertical + per-tenure + per-RFM + per-seasonality + per-loyalty-tier + per-subscription-status + per-acquisition-source + per-first-order-SKU) + model ensemble decision (per-BG/NBD + per-Pareto/NBD + per-Gamma-Gamma + per-MBG/NBD + per-CDNOW + per-LSTM + per-Transformer + per-Convolutional Sequence-to-Sequence with stacking / blending weights + cross-validation evidence + MLflow + Weights & Biases + CometML run pointer) + per-customer CLV score + per-customer confidence interval (Bayesian + frequentist + bootstrap) + per-customer confidence tier + per-customer embedding pointer + per-customer RFM record + per-cohort recalibration record (Platt + isotonic + temperature + Bayesian binning + ECE + reliability diagram + Brier score + recalibration cadence) + per-customer runtime cache pointer + per-customer explainability record (SHAP + LIME + feature importance + counterfactual + attribution graph + multi-LLM narrative under per-vendor zero-retention) + per-anchor Gate decision with evidence (CCPA opt-out posture + GDPR Article 22 disclosure + EU AI Act Article 50 disclosure + EU AI Act Annex III posture + FCRA permissible purpose + adverse-action notice posture + ECOA disparate-impact audit + Fair Housing audit + FASB ASC 606 reconciliation evidence + SEC Reg G reconciliation evidence) + per-vendor LLM zero-retention verification + cross-skill handoff record + FBC feedback loop record. Storage on AWS S3 Object Lock + Azure Blob immutable + Google Cloud Storage Bucket Lock + Wasabi compliance WORM. Retention stacks (longest applicable wins): 7-year FTC substantiation + 7-year IRS + 7-year FDD + per-state franchise registration + GDPR Article 30 records of processing + CCPA 12-month look-back + EU AI Act Article 12 record-keeping + SOX Section 802 7-year + SOC 2 CC7 / CC8. DSAR export shipped on demand to satisfy CCPA Right to Know + CPRA right to correct + GDPR Article 15 right of access + GDPR Article 17 right to erasure + 17-state DSAR. End-to-end replay rewinds Cohort + Recalibrate + Gate + DSAR export with confidence tier and explainability at every stage. Sibling handoffs flow into per-member monthly CLV, per-member next-best-action, churn-prediction per-subscriber sibling build-pillar, tier-transition timing, save-flow propensity scoring, cross-location offer coordination sibling build-pillar, foot-traffic integration sibling build-pillar, cross-touchpoint identity resolution sibling build-pillar, attribution event emission, the list-segmentation sibling skill (per-LTV-quintile is one of the 14 standing segmentation strategies — CLV is the substrate that feeds it), and the monthly-executive-summary-drafting build-pillar on the rollup-reporting agent (CLV feeds customer / loyalty section variance commentary).

Engage Completions on the customer-data-graph bundle

The Cohort + Recalibrate + Gate + Audit four-skill bundle ships as the orchestration layer above your existing vendor CLV + embedding store + runtime cache + streaming compute surface. CCPA right to opt out of automated decision-making + EU AI Act Article 22 + Annex III + FCRA + ECOA + Fair Housing + Brier score calibration discipline + FASB ASC 606 + SEC Reg G + NIST AI RMF anchors are preserved in every per-customer audit record. Tier 1 AI Readiness Assessment scopes the bundle in two to three weeks; Tier 3 Fractional CMO with AI Swarm operates the bundle end-to-end.

Tier 1 AI Readiness Assessment Tier 3 Fractional CMO with AI Swarm Take the 3-question quiz