Event-Driven AI in Commerce: How Tanqory Interprets Signals in Real Time

Event-Driven AI in Commerce: How Tanqory Interprets Signals in Real Time
Research blog — architecture, methods, and evidence for Tanqory's event-driven AI stack

Executive Summary

Commerce is increasingly defined by what happens between events: a price change, a stockout, a payment retry, a search reformulation, or a social spike. Tanqory treats AI not as an offline modeling exercise but as an event-driven control system. Our stack ingests heterogeneous signals, attaches context and consent, infers causality where possible, and routes actions with latency, fairness, and trust guardrails. This post details the Event Signal Fabric (ESF) and Decision Orchestration Layer (DOL) that power Tanqory's real-time interpretation of demand, risk, and opportunity.

Highlights:

• Event-first data plane with schema contracts, residency, and consent baked in.
• Causal uplift and synthetic controls to separate signal from noise (Athey and Imbens, 2016; Abadie et al., 2010).
• Contextual bandits and constrained planners for decisions under latency and fairness budgets (Agarwal et al., 2014; Patil et al., 2021).
• Control-inspired regulators smooth volatility and prevent overreaction during spikes (Deng et al., 2013 for variance reduction; Box et al., 2015 for time-series stability).
• Observability, privacy, and auditability as first-class features to sustain trust (Dwork and Roth, 2014; Kohavi et al., 2020).

1. Why Event-Driven AI Matters in Commerce

1.1 Limitations of Batch-First AI

Batch pipelines deliver models days or weeks late, missing transient intent and operational shocks. This leads to:

• Latency between signal and action; demand fades before interventions land.
• Correlation traps: weekly aggregates obscure causal levers (Pearl, 2009).
• Reliability gaps: models drift quietly without on-path guardrails (Kohavi et al., 2020).
• Trust erosion: customers see stale promises and irrelevant offers (Gefen et al., 2003).

1.2 Event-Driven Benefits

• React within minutes to search reformulations, payment retries, or carrier incidents.
• Preserve causal interpretability by capturing temporal ordering and treatment timing.
• Maintain fairness and safety with on-path checks rather than offline audits only.
• Reduce decision latency so merchandising, pricing, and support stay aligned with real conditions.

2. Design Principles

1. Event-first: every meaningful state change is an event with schema, consent, and residency flags.
2. Causal preference: uplift models, synthetic controls, and natural experiments over pure correlation (Athey and Imbens, 2016; Abadie et al., 2010).
3. Constraint-aware: fairness, latency, privacy, and cost budgets enforced at decision time (Patil et al., 2021; Dwork and Roth, 2014).
4. Separation of reasoning and execution: ESF senses and reasons; DOL acts with rollbacks.
5. Human-in-the-loop for high impact: operators review high-risk changes; explanations accompany recommendations (Lundberg and Lee, 2017).
6. Observability and audit: full trace from event to action to outcome, exportable for regulators and partners.
7. Multi-region readiness: per-region routing, localization, and fairness thresholds to respect heterogeneity.

3. Architecture Overview

3.1 Event Signal Fabric (ESF)

• Ingest and Contracts: Kafka-like bus; schemas enforced with registry; consent/residency tags on every event.
• Signal Normalization: deduplication, watermarking for late events (Kleppmann, 2017); enrichment with region, device, cohort.
• Feature and Signal Store: online/offline parity; freshness SLAs (<15 minutes for critical behavioral signals); lineage captured for audit.
• Quality Monitors: null thresholds, distribution drift (Kullback and Leibler, 1951), volume anomalies.

3.2 Decision Orchestration Layer (DOL)

• Causal Services: uplift models, synthetic controls, IV-based estimators for natural experiments.
• Predictive Models: gradient-boosted trees for tabular reliability (Chen and Guestrin, 2016); lightweight Transformers for sequence signals under latency caps (Vaswani et al., 2017).
• Policy Engine: contextual bandits with fairness/latency constraints (Agarwal et al., 2014; Patil et al., 2021), planners for multi-step actions (pricing + assortment + promise shaping).
• Safety Layer: consent and residency enforcement, bias checks, p95/p99 latency guards.
• Action Connectors: merchandising, pricing, CMS, supply chain orchestration, support tooling.

4. Signal Taxonomy and Engineering

4.1 Behavioral and Intent Signals

• Search queries and reformulation chains; query entropy as focus proxy (Huang et al., 2020).
• Clickstreams, scroll depth, dwell, backtracks; cart edits and wishlist dynamics.
• Session context: device, locale, referral, bandwidth.

4.2 Commerce and Operational Signals

• Inventory and promise accuracy; substitutions; node congestion.
• Pricing changes, discounts, fees, taxes; elasticity probes.
• Payment retries, decline codes, fraud scores.
• Support tickets, CSAT, refund reasons; trust posture (Gefen et al., 2003).

4.3 External and Macro Signals

• Exchange rates, inflation prints, holiday calendars.
• Social and search trend indices (Choi and Varian, 2012); caution against big-data traps (Lazer et al., 2014).
• Weather and mobility where relevant to logistics.

4.4 Data Quality Signals

• Freshness coverage per feature; alert <90% refreshed.
• Volume anomalies; schema rejects; field-level drift (KL).
• Lineage completeness for audit and reproducibility.

5. Causal and Predictive Modeling

5.1 Causal Uplift and Experiments

• Doubly robust uplift for promos and placements (Athey and Imbens, 2016).
• Synthetic controls for regions where experiments are infeasible (Abadie et al., 2010).
• Sequential tests to avoid peeking bias (Johari et al., 2017).
• Variance reduction with CUPED to stabilize detection during spikes (Deng et al., 2013).

5.2 Forecasting and Nowcasting

• Time-series: SARIMA/Prophet-like for medium horizon; Box et al. (2015) guidance on control and seasonality.
• Sequence/transformer: intent and traffic nowcasting under p95/p99 budgets (Vaswani et al., 2017).
• Hierarchical models: borrow strength across regions/categories; manage sparsity.
• Anomaly detection: seasonal hybrid ESD; autoencoders for rare events.

5.3 Explainability and Trust

• SHAP on boosted trees for operator insight (Lundberg and Lee, 2017).
• Counterfactual prompts: what if price drops 5%? what if promise improves 1 day?
• Influence shift alerts when feature contributions drift unexpectedly.

6. Decision Policies and Constraints

6.1 Bandits with Guardrails

• Contextual bandits allocate exposure for content, promos, and layouts.
• Fairness caps: halt if uplift gap >5pp across protected cohorts (Patil et al., 2021).
• Latency guard: p95 <100ms, p99 <150ms; auto-fallback if breached.
• Exploration tuned by cohort volatility; suppressed for low-signal or at-risk segments.

6.2 Planners for Multi-Step Actions

• Joint decisions on pricing + assortment + promise shaping; constraints on margin, capacity, and trust.
• Linear/quadratic solvers enforce budget, inventory, and SLA limits.
• Safety checks: consent, residency, and policy filters (e.g., age/legal constraints).

6.3 Control-Inspired Stabilizers

• PID-like regulators smooth exposure during spikes, reducing KPI whiplash.
• Apply dampening when uplift variance spikes or when volume exceeds forecast bands (Box et al., 2015).
• Avoid overreaction to noisy events (Lazer et al., 2014).

7. From Signals to Actions

7.1 Assortment and Category Bets

• Detect micro-trends from search reformulations and wishlists; validate with synthetic controls.
• Recommend onboarding suppliers or reallocating inventory; tie to promise reliability to avoid stockout-driven churn (Zheng and Federgruen, 2020).

7.2 Pricing and Promotion

• Elasticity from causal graphs and experiments; retargeting insights (Lambrecht and Tucker, 2013) inform personalization caps.
• Bandits allocate promos with fairness and latency constraints; auto-rollback on bias or latency breach.

7.3 Experience and Localization

• Surface localized copy when intent signals imply mismatch; adjust navigation for friction topics.
• Keep render and decision latency under guardrails; fallbacks ready.

7.4 Payments and Trust

• Monitor issuer-specific declines; trigger alternative payment offers; track trust posture via disputes and CSAT (Gefen et al., 2003).
• Fraud and risk scores feed exposure throttles; governance on data use (Dwork and Roth, 2014).

7.5 Supply and Promise Shaping

• Link demand signals to fulfillment constraints; adjust exposure and promises based on congestion forecasts.
• Route through orchestration engine to reroute at-risk corridors; preserve promise accuracy (a driver of conversion and retention).

8. Observability, Reliability, and Governance

8.1 Metrics

• Decision health: latency, timeout, fallback, fairness gaps.
• Data health: freshness, nulls, schema rejects, drift.
• Model health: calibration, uplift stability, OOD detection.
• Business: revenue per session, conversion, cancellation, promise accuracy.

8.2 Tracing and Audit

• Trace IDs across ingest → features → decision → render.
• Log policy version, model hash, feature snapshot, cohort tags.
• Replay harness for audits; exportable logs for regulators.

8.3 Governance and Compliance

• Experiment registry (Kohavi et al., 2020); hypotheses, guardrails, rollbacks.
• DPIA for new signals; region-bound storage; consent and residency checks on-path (Dwork and Roth, 2014).
• Access controls; change management for high-risk policies.

9. Multi-Region Adaptation

• Per-region parameterization for latency budgets, fairness thresholds, and consent rules.
• Localization of tokenization, sentiment, and holiday calendars.
• Hierarchical priors and transfer learning to mitigate sparsity; conservative exploration in new markets.
• Blast-radius isolation: region-scoped rollouts and kill switches.

10. Experimentation Framework

• Pre-registration of hypotheses and metrics; MDE defined.
• Sequential testing to control peeking bias (Johari et al., 2017).
• CUPED for variance reduction (Deng et al., 2013).
• Holdouts for drift monitoring; bias and latency guardrails with auto-rollback.

11. Performance and Cost Engineering

• Latency budgets: ingest (10-20ms), feature read (20-30ms), decision (25-40ms), render (20-30ms).
• Caching: hot features and embeddings; rule fallbacks for low bandwidth.
• Cost dashboards: per-model and per-feature cost per 1,000 decisions; prune unused artifacts.
• Chaos drills and replay harness for surge traffic.

12. Security, Privacy, and Ethics

• Least privilege access; secrets vaulted; mTLS between services.
• PII hashing; redaction in logs; residency enforcement.
• Fairness scans and exposure caps; review boards for high-impact changes.
• Avoid manipulative patterns; user autonomy and opt-outs (Thaler and Sunstein, 2009).
• Transparency: operator explanations; user notices where required.

13. Case Studies (Illustrative, Evidence-Aligned)

• EU Fashion: search reformulation spikes + synthetic controls signaled genuine trend; localized copy + controlled promos lifted conversion 6.8%; aligns with intent-cue literature (Huang et al., 2020).
• US Electronics: issuer-specific declines detected; alternative payments restored 5.9% of at-risk orders; supports trust pathways (Gefen et al., 2003).
• APAC Beauty: social trend acceleration validated with Google Trends-style nowcasting (Choi and Varian, 2012); inventory reallocation reduced stockouts and markdowns.
• LATAM Home: promise accuracy uplift after carrier mix change improved conversion 4.3%; consistent with reliability effects on trust.
• MEA Cross-Border: fairness-capped bandits kept uplift gaps <4pp while improving CVR 5.1% (Patil et al., 2021).

14. Data Products and Interfaces

• Decision APIs: ranked offers, layouts, and promises with metadata (policy version, fairness status, confidence).
• Analyst workspace: semantic layers to reduce metric drift; notebooks for uplift, drift, synthetic controls.
• Partner reporting: exposure transparency, attribution with confidence bands, residency and consent summaries.
• Signal health index: freshness and quality score per feature; alerts to suppress risky signals.

15. Operator Workflow

1. Intake: signal/use-case proposal with DPIA checklist.
2. Design: causal graph, metrics, guardrails, rollout plan.
3. Build: implement features/models; tests for schema, freshness, drift.
4. Validate: offline eval, simulation, privacy/fairness review.
5. Launch: shadow → gated ramp; observability on.
6. Operate: monitor metrics; trigger playbooks on anomalies.
7. Review: postmortems; update thresholds and playbooks.

16. Technical SLAs and Playbooks

• Ingest availability 99.9%; schema reject <0.1% after launch; freshness <5 minutes for critical signals.
• Feature store p95 read <30ms; alerts at 90% freshness.
• Decision services p95 <100ms, p99 <150ms; timeout budget <0.5%.
• Playbooks: data quality (quarantine → backfill → replay), drift (fallback → recalibrate → shadow), fairness (halt → analyze gaps → adjust constraints), latency (lightweight policy → scale infra → profile).

17. KPI Math

• Revenue per session, contribution margin, promise accuracy, cancellation rate.
• Fairness gap = max uplift - min uplift across protected cohorts; halt if >5pp for >1 hour.
• Drift score: KL vs 7-day baseline; investigate >0.08; rollback >0.12.
• Fallback hit rate: fallback/total decisions; leading indicator of instability.

18. Limitations and Open Questions

• Unobserved confounding: natural experiments are uneven; causal claims require humility (Pearl, 2009).
• Data sparsity: low-signal cohorts need conservative exploration and longer tests.
• Attribution: multi-touch journeys blur impact; ongoing causal multitouch work.
• Adversarial noise: bot traffic and spam require continuous detection.
• Human factors: operator interpretation of explanations varies; UX research needed.
• Sustainability: energy/carbon not yet first-class in optimization; roadmap includes multi-objective tuning.

19. Future Research Directions

• Adaptive exploration blending Bayesian bandits with fairness constraints.
• Federated/on-device learning to reduce data movement and latency.
• Causal discovery that tolerates feedback loops; stability checks across regions.
• Carbon-aware decisioning: integrate energy signals into reward functions.
• Explainability UX studies to reduce cognitive load for operators.
• Trust integration: tighter coupling between fraud, content safety, and experience policies.

20a. KPI Math and Thresholds

• Revenue per session: revenue/sessions, tracked by cohort; paired with variance bands.
• Contribution margin: revenue minus variable costs (fulfillment, payments, returns); primary for pricing and exposure.
• Promise accuracy: 1 - |actual - promised|/promised window; penalizes early and late equally.
• Cancellation rate: cancellations/orders; segmented by promise accuracy deciles.
• Fallback hit rate: fallback/total decisions; leading indicator of instability.
• Fairness gap: max uplift - min uplift across protected cohorts; auto-halt if >5pp for >1 hour.
• Drift score: KL vs 7-day baseline (Kullback and Leibler, 1951); investigate >0.08; rollback >0.12.
• Freshness: percent of critical features refreshed within SLA; alert if <90%.

20b. Technical SLAs and Capacity

• Ingest: 99.9% availability; schema reject <0.1% post-launch; freshness <5 minutes for critical signals.
• Feature store: p95 read <30ms; freshness alerts at 90%.
• Decision services: p95 <100ms, p99 <150ms; timeout budget <0.5%.
• Connectors: idempotent writes; DLQs monitored; retries with backoff.
• Capacity drills: replay harness for surge; chaos tests for upstream outages; per-region scaling.
• Cost controls: per-model cost alerts; auto-park unused models/features.

20c. Operator FAQ (Condensed)

• Is a spike real or noise?
  Confirm with synthetic controls, placebo tests, and multi-signal corroboration (behavioral + external).
• What if data is sparse?
  Use hierarchical priors, conservative exploration, longer tests, and qualitative checks.
• Privacy/consent enforcement?
  Consent/residency flags on events, region-bound storage, DPIA for new signals, auditable logs.
• How fast can we act?
  Shadow within hours; gated ramp within a day if guardrails green; latency/fallback keep UX safe.
• What stops harmful policies?
  Fairness/latency/privacy guardrails, experiment registry, review boards, auto-rollbacks, kill switches.

21. Partner and Merchant Reporting

• Exposure transparency: why SKUs/experiences were promoted or throttled; includes signal health and promise reliability.
• Action recommendations: localized pricing, copy, and fulfillment adjustments for detected friction.
• Attribution clarity: uplift with confidence bands; distinguish platform vs merchant effects.
• Feedback loops: partners can flag false positives or local context; labels feed retraining.
• Compliance summaries: residency routing, consent coverage, fairness parity for auditors.

22. Human + Machine Collaboration

• Override mechanisms: operators can pause/adjust policies when qualitative evidence conflicts; overrides logged for learning.
• Cultural context panels: regional experts annotate anomalies (holidays, sentiment nuance); annotations become features.
• Training: onboarding for PM/DS/Eng on causal reasoning, fairness, and latency budgets.
• Decision review boards: monthly reviews of high-impact experiments and policy changes.

23. Future Research Directions

• Adaptive exploration blending Bayesian bandits with fairness and latency constraints; formalize stopping rules under multi-objective rewards.
• Federated and on-device learning to reduce data movement and improve privacy/latency; measure personalization trade-offs.
• Causal discovery that tolerates feedback loops; stability checks across regions and seasons.
• Carbon-aware decisioning: integrate energy signals into reward functions; study user tolerance for greener options.
• Explainability UX: research operator comprehension of SHAP/counterfactuals; design low-friction summaries.
• Trust integration: couple fraud, content safety, and experience policies to avoid cross-domain regressions.

20. Conclusion

Event-driven AI turns commerce into a responsive, governed system. By capturing signals with context, applying causal and constrained decisioning, and enforcing on-path guardrails, Tanqory reduces decision latency while protecting fairness, privacy, and trust. The ESF and DOL make AI an operational capability, not a batch report.