LedgerGuard.

Operational evidence

A workflow is not production-minded until it can be measured, paused, retried, and audited.

These are real aggregates from invoices the pipeline processed—not a fixture total. The page exposes throughput, outcome mix, backlog, exception frequency, integration health, per-invoice latency, model cost, alerts, and missing capabilities.
5 decided invoices 53 audit events · $0.058 cost

Live operating snapshot

What the pipeline has actually done

5
Invoices decided
53
Audit events recorded
35.5s
Total processing time
$0.058
Total model cost

Invoices by state

1 blocked
1 duplicate hold
2 exception review
1 ready for approval

5 total invoice row(s) — 0 are seeded historical/duplicate-detection data that never entered the decision pipeline (see /queue).

Duplicate detection & approval backlog

1 / 1
Confirmed duplicate holds / invoices with any candidate signal
0%
Human correction rate (decided invoices with a field correction)
3
Open in the approval backlog right now
Invoice Outcome Age
SPH-40917 Exception review 977.3h
CSS-72104 Blocked 977.3h
PGL-61144 Exception review 977.3h

Exception frequency by control

Unit-price tolerance — Emergency compressor unit replacement 1
Quantity tolerance — HVAC technician labor, emergency repair 1
Normalized duplicate-identity match 1
Remittance bank-detail comparison 1
Unmatched invoice line 1
Embedded-instruction screening 1

Integration health

0
Draft bills created (accounting sandbox)
0
Accounting-write failures
0
Jobs retrying (transient failure)
0
Jobs dead-lettered (permanent failure)

Per-invoice summary

Invoice Outcome Events Latency Cost
APC-88213 Duplicate hold 9 5.6s $0.0100
SPH-40917 Exception review 11 8.3s $0.0136
PGL-61144 Exception review 11 8.1s $0.0118
BJS-55821 Ready for approval 12 8.4s $0.0134
CSS-72104 Blocked 10 4.9s $0.0095

Alerts — live in v1

  • Critical bank-detail change detected
    Fires the moment a bank_detail_change control fails — see Scenario 4.
  • Duplicate accounting write attempted
    Fires if a replayed submissionId or idempotency key ever reaches the accounting integration.
  • OCR or model cost exceeds daily limit
    Wired to the same race-safe daily spend-cap ledger used across the portfolio.

Production-documented, not built in v1

  • · Accounting-integration authentication failure
  • · Extraction-failure-rate threshold exceeded
  • · Monetary-correction-rate increase
  • · Exception-backlog SLA breach
  • · Workflow stuck beyond expected duration
  • · Evaluation regression crosses a release threshold

Operational controls

  • Pause accounting writes independently from extraction.
  • Disable a failing OCR or model provider.
  • Reprocess a workflow from its last successful stage.
  • Version extraction prompts, schemas, rules, and models — display the version behind every decision.
  • Replay dead-lettered jobs after remediation.
  • Require dual approval to change critical controls (documented; single-operator in this demo).

Not yet built on this dashboard

  • Extraction confidence trended by field (needs a rolling baseline this project doesn't keep yet).
  • Supplier and document-layout drift detection (same — needs historical trend data, not a single snapshot).

Built by Ariel

Ariel Magalso
AI Engineer
Philippines · Remote

Operations are part of the engineering proof.

The dashboard demonstrates auditability, cost visibility, backlog ownership, integration health, retry strategy, alerts, and an explicit list of capabilities that are documented but not yet built.

Open to opportunities

Need AI automation that can explain itself?

I build measurable AI-assisted workflows with deterministic safeguards, visible evaluation, and human review where the risk demands it.

Contact Ariel