B2B Data and AI Services Firm / Innodata
Challenge
A B2B data and AI services firm ran paid media across four ad platforms, Google Analytics, Pardot, and Salesforce with no single source of truth. Each system reported different numbers. Reconciliation happened manually, and campaign problems surfaced days or weeks after the budget was already spent. A second problem was trust: for a marketing leader who already did not trust the numbers, an AI system that returned a plausible-sounding but invented metric would be a liability. Cross-channel attribution added a third layer of complexity, since connecting a lead back to the campaign that produced it requires confidence tiering, and a confident attribution number with no caveat drives bad budget decisions.
A B2B data and AI services firm ran paid media across four ad platforms, Google Analytics, Pardot, and Salesforce with no single source of truth. Each system reported different numbers. Reconciliation happened manually, and campaign problems surfaced days or weeks after the budget was already spent. A second problem was trust: for a marketing leader who already did not trust the numbers, an AI system that returned a plausible-sounding but invented metric would be a liability. Cross-channel attribution added a third layer of complexity, since connecting a lead back to the campaign that produced it requires confidence tiering, and a confident attribution number with no caveat drives bad budget decisions.
Solution
A five-agent intelligence system runs on Wippy, an open-source actor-model runtime, backed by a single PostgreSQL warehouse on AWS. The core architectural constraint: every numerical answer must be produced by a verified SQL query against the warehouse, with that query logged and inspectable. The warehouse uses three stages: raw platform data stored untouched, normalized to a shared vocabulary, then served exclusively to agents from the normalized layer. Attribution windows are re-pulled automatically to handle conversion data that restates over time. Cross-channel attribution uses a three-tier confidence model with the tier disclosed on every answer. Anomaly detection rules are stored as editable configuration the team adjusts without a developer. Read-only access is enforced at three independent layers: platform credentials, SQL tool constraints, and agent-level instructions.
A five-agent intelligence system runs on Wippy, an open-source actor-model runtime, backed by a single PostgreSQL warehouse on AWS. The core architectural constraint: every numerical answer must be produced by a verified SQL query against the warehouse, with that query logged and inspectable. The warehouse uses three stages: raw platform data stored untouched, normalized to a shared vocabulary, then served exclusively to agents from the normalized layer. Attribution windows are re-pulled automatically to handle conversion data that restates over time. Cross-channel attribution uses a three-tier confidence model with the tier disclosed on every answer. Anomaly detection rules are stored as editable configuration the team adjusts without a developer. Read-only access is enforced at three independent layers: platform credentials, SQL tool constraints, and agent-level instructions.
Results
The engagement produced a complete architecture and system-design specification for the five-agent system, a normalized PostgreSQL warehouse design with a preserved raw layer, isolated per-platform connectors for all six data sources, a three-tier attribution model with confidence disclosure, a rules-as-data anomaly engine seeded from the team's own decision frameworks, and read-only enforcement at three independent layers. The client owns their data, agent configurations and thresholds, decision rules, and every report and alert the system produces. Thresholds are adjustable without any dependency on Spiral Scout for routine changes.
The engagement produced a complete architecture and system-design specification for the five-agent system, a normalized PostgreSQL warehouse design with a preserved raw layer, isolated per-platform connectors for all six data sources, a three-tier attribution model with confidence disclosure, a rules-as-data anomaly engine seeded from the team's own decision frameworks, and read-only enforcement at three independent layers. The client owns their data, agent configurations and thresholds, decision rules, and every report and alert the system produces. Thresholds are adjustable without any dependency on Spiral Scout for routine changes.
