ServiceLink EXOS | Analytics
Synthetic demonstration data — structurally faithful to production OpenTelemetry. Every metric is a live ClickHouse query.
Process Mining · Product Analytics · Session Replay

Find the operator workflows worth automating.
Prove which ones aren't.

EXOS turns real OpenTelemetry traces into operational process mining, product analytics, and session replay — all on ClickHouse, all SQL-native. It surfaces the high-frequency, high-determinism workflows that are ready to automate, and the bespoke ones that genuinely aren't.

780K+ OpenTelemetry spans
410K+ browser RUM sessions
SQL every metric is a query
ClickHouse on GCP us-central1
Automation candidate /api/task-detail
status_request 1,284 emails · email channel
Inbound state → extracted entities (confidence, not key-presence)
order.id 0.96
borrower 0.91
property_address 0.88
loan_amount 0.74

Response template concentration T1 · 94%
Median edit-distance from template ≈ 0
Automate 0.92
escalation → bespoke, dispersed 0.11

Automatability is earned from response determinism given the extracted inbound state — not asserted from a bare received → sent sequence. Illustrative readout.

How it works

Real traces in. SQL-native analysis out.

No black-box pipeline and no fabricated datasets — the data is synthetic, but the path from raw span to insight is the real one you'd run in production.

01
Real OpenTelemetry traces

Browser RUM plus backend spans, modeled to OTel semantics — resource attributes (telemetry.sdk.*, service.version, deployment.environment), span kinds, realistic latencies, and clustered error skew.

02
ClickHouse

Every span lands in a columnar store on GCP. There's no pre-aggregated cache standing in for the data — each panel hits otel_traces directly. Hover any metric on the dashboard to see the exact query behind it.

03
SQL-native mining

Process mining, lifecycle funnels, and anomaly detection (a multivariate z-score at 2.5σ) all run as SQL — auditable, reproducible, and portable. The analysis is the query, not a vendor's secret.

Synthetic data, real analysis

The traces in this demo are generated, not captured from production — so the numbers will differ from a live deployment, and nothing here is a customer's data. What is not faked is the analysis: there are no Math.random() fallbacks and no fabricated result sets — every figure is a live ClickHouse query you can read. The synthetic corpus is modeled to be structurally indistinguishable from production OpenTelemetry (spec-faithful resource attributes, span kinds, latency models, and error distributions) so the mining and analytics features have something legible and realistic to operate on.