← Index
← Back to selected work
Case 02 / Lumen Technologies · 2025–

RADAR.
Compliance that
survives a departure.

A six-tier regulatory operations platform on GCP that encodes monthly PUC filing knowledge into a structured, queryable form — so the next time the report owner leaves, the system doesn't have to be rebuilt from scratch.

RoleDesigner · builder · owner
Coverage8 US states · monthly filings
StackBigQuery · Vertex AI · BQML
GovernanceOneTrust 565 (approved)
01 · The motivation

Lumen files monthly PUC regulatory reports across 8 states — Oregon, Idaho, Minnesota, Nebraska, New Mexico, Montana, Iowa, Wyoming — each carrying monetary penalties when service-quality thresholds are missed.

Three problems kept compounding. Single-person dependency: when the Oregon owner left, the SQL, exception rules, and threshold knowledge weren't fully transferable. Context-free monitoring: existing tools could tell you a value dropped 5.6% — they couldn't tell you whether that crossed a regulatory threshold. Rule drift: report-specific logic lived in people's heads, so new operators ran reports without institutional knowledge and downstream decisions used wrong numbers.

RADAR makes it impossible to lose this knowledge again.

02 · The Report Genome

The central idea: every recurring report gets a Genome — a structured, queryable knowledge profile that captures what the report is, what it depends on, and what makes it pass or fail.

Each Genome encodes purpose and regulatory obligation, upstream ETL freshness SLAs, 90-day rolling baselines per metric, regulatory thresholds sourced from published PUC rules, cross-jurisdiction Pearson correlations with neighbour states, failure history with recovery playbooks, cost and bytes scanned per run, and the blast radius — which other reports break if this one fails.

03 · Six-tier architecture

Hosted on GCP (prj-mmopsrpt-prod-001) with BigQuery as the analytical core. Five deterministic analytical methods run on every execution, with a thin Vertex AI layer above them as a translation surface only.

Tier 1
Report Genome
Structured knowledge profile per report. Purpose, dependencies, baselines, thresholds, blast radius.
Tier 2
Cross-jurisdiction correlation
SQL JOIN against correlation profiles. Separates data-quality issues from regional events.
Tier 3
Threshold-aware classification
CASE statement: safe / approaching / crossing / recovering — based on regulatory proximity.
Tier 4
Counterfactual correction
Filtered aggregations ranked by marginal impact. Minimum record set whose correction restores compliance.
Tier 5
Predictive trajectory
BigQuery ML ARIMA_PLUS model projects metrics to filing deadline. Flags expected crossings early.
Tier 6
Dynamic penalty escalation
Arithmetic over the penalty schedule. Live financial exposure including per-day compounding.
04 · LLMs as a thin translation layer

Vertex AI (Gemini) sits above the deterministic core, never inside it. It receives structured analytical outputs and generates natural-language advisories, filing narratives, root-cause hypotheses, and conversational answers.

Gemini does not analyse. If you removed it tomorrow, RADAR continues to function using template text. A second LLM provider is wired up as a backup model. This separation is deliberate — analysis is deterministic and auditable; the LLM is a translator only.

05 · Delivery & governance

Microsoft Teams alerts, email digests, QlikSense dashboards. Full audit trail in Cloud Logging. Submitted through Lumen's AI governance process as OneTrust Project 565 — approved.

15–30→1
Min of dashboard hunting → one grounded conversational question
1–2hr→review
Filing narrative drafting → review-and-approve
8
US states · institutional knowledge encoded
94%
BigQuery bytes scanned reduction (sister project)
“Compliance knowledge is now institutional — encoded in the Report Genome — rather than held by any individual.”— RADAR design memo
Next case
Grab — real-time IoT pipeline across SE Asia