Reliability · continuous improvement

Tollscopic treats tolling reliability like a software and data-quality problem.

Observe the live system, capture edge cases, replay historical evidence, regression-test changes, stage releases, and monitor the result. The improvement loop is the product.

01 · The improvement loop

Five stages. The output of each becomes the input of the next.

This is the page where Tollscopic sounds like a serious AI company without using hype: observe production, capture what is hard, replay safely, test rigorously, release in stages.

01 Observe 02 Capture 03 Replay 04 Test 05 Release VERSION TRACEABILITY · evidence chain stays intact
01 · OBSERVE

Device health, image quality, evidence flow, queue depth, latency, model quality.

02 · CAPTURE

Edge cases, drift events, anomalies, low-confidence transactions, missing reads.

03 · REPLAY

Re-run historical evidence against candidate model or configuration changes.

04 · TEST

Regression-test changes against a held-out replay corpus. Promote only on improvement.

05 · RELEASE

Stage the change to a subset. Monitor outcomes. Keep version traceability.

02 · Observability layers

The system watches the pipeline, not just servers.

A green dashboard that says 'all servers up' is not enough for tolling. The signals that matter are the ones that predict KPI failure before it happens.

01 · LAYER
Roadside · device
camera heartbeat
image quality
frame rate
PTZ events
lens condition
02 · LAYER
Edge compute
CPU/GPU load
queue depth
evidence buffer
failover state
OTA version
03 · LAYER
Cloud · ingestion
event throughput
ingestion latency
schema validation
duplicate rate
04 · LAYER
Transaction · quality
confidence distribution
classification distribution
missing-evidence rate
low-conf rate
05 · LAYER
Back-office · delivery
publish latency
ack rate
retry depth
reconciliation gap
03 · Drift detection

What happens before a KPI failure becomes obvious.

A camera slowly loses alignment. Image confidence falls. A classification distribution shifts. A queue grows during WAN degradation. Tollscopic monitors for these signals so they can be investigated early.

01
Image quality drift

Per-camera confidence trending down. Brightness/contrast shifts. Frequent PTZ events.

02
Classification drift

Class distribution shifting from baseline without a corresponding traffic-pattern change.

03
Evidence-volume drift

A device producing fewer events than expected. Queue growth. Latency creeping up.

04 · Resilience

Imperfect networks. Imperfect weather. Still publishing cleanly.

The roadside operates in real conditions. The system preserves evidence locally when needed, publishes carefully when connectivity returns, and avoids duplicate or lost records through explicit state handling.

01
Local buffering

Evidence is persisted at the edge during WAN degradation. The roadside keeps observing.

02
Duplicate-safe publishing

Idempotency keys and explicit state tracking. Reconnects do not produce duplicate transactions.

03
Hot failover

Redundant Jetson Orin pair per cabinet. Edge compute survives a single-unit failure.

04
Cloud redundancy

Multi-AZ cloud, event-driven processing, explicit recovery patterns for transient faults.

Reliability is what stops a tolling system from going sideways at month-end.

If your current vendor's reliability story is about response-time SLAs and maintenance schedules, this page is the contrast.