Recruiter case study · 5 minute read
01 / 010
Accounts-payable teams repeatedly copy invoice data, verify arithmetic, reconcile purchase orders and receipts, check for duplicates and supplier-detail changes, and decide who must review each exception. Delays hold up suppliers and operations; a confident wrong answer can create a duplicate payment, misclassification, fraud exposure, or real financial loss.
02 / 010
The workflow is high-volume and repetitive, but it still contains judgment points that should remain accountable. Automating document preparation and deterministic checks can shorten review cycles, surface risky exceptions earlier, preserve a consistent audit trail, and let finance teams spend more time resolving discrepancies instead of re-keying routine data.
03 / 010
LedgerGuard uses AI where invoice layouts are variable: locating and extracting candidate fields and drafting evidence-linked explanations. Deterministic code handles money, identity, duplicate rules, tolerances, routing, permissions, and idempotency. A human reviews the prepared work, and the model never decides that an invoice is financially safe.
04 / 010
A suitable implementation can reduce manual preparation, standardize controls across teams and locations, make exception queues easier to prioritize, improve traceability, and create a reusable integration layer around existing finance systems. These are intended workflow benefits, not measured customer outcomes from this fictional portfolio demonstration.
05 / 010
LedgerGuard is most relevant to organizations processing meaningful invoice volume across multiple suppliers, locations, purchase orders, receipts, or approval routes. Potential fits include facilities and property management, construction, hospitality, multi-site retail, manufacturing, logistics, healthcare administration, and professional-services companies with lean finance teams.
06 / 010
The workflow can be connected through APIs, webhooks, and system-specific adapters. It can receive documents and reference data, return status and evidence, notify reviewers, and prepare draft accounting records. The safety boundary remains unchanged: integrations may prepare work, but payment authorization stays outside the AI workflow.
07 / 010
I independently designed and implemented the FastAPI application, Claude extraction pipeline, evidence alignment, PostgreSQL data model, matching and control engine, five-scenario workbench, approval queue, evaluation harness, operations dashboard, sandbox accounting adapter, and deployment.
08 / 010
Every extracted field must point to document evidence. Monetary values are recalculated with decimal-safe code. Bank-detail changes are blocked, duplicate identity survives renamed files, embedded instructions remain untrusted document text, and every proposed external write carries an idempotency key.
09 / 010
The evaluation runner sends labeled fictional invoices through the actual extraction, alignment, matching, and decision pipeline. Held-out and development splits are reported separately, failures remain visible, and critical false clearances are measured independently from field-level extraction errors.
010 / 010
This is a portfolio demonstration, not a customer deployment. All entities and financial data are fictional. The public workflow intentionally stops before executing an accounting write or payment, and the evaluation set is still small compared with a production document population.
Industry fit
These examples describe organizations that could benefit from this architecture. They are illustrative applications, not current LedgerGuard customers.
Recurring maintenance, security, cleaning, and repair invoices routed across properties and cost centers.
Subcontractor and materials invoices reconciled with project purchase orders, receipts, and tolerances.
Distributed invoice intake standardized across locations while local managers retain approval authority.
High-volume supplier charges checked against orders, deliveries, and approved commercial terms.
Non-clinical supplier invoices prepared for controlled review with evidence and audit history.
Growing organizations reduce repetitive AP preparation without removing human financial oversight.
Integration-ready design
LedgerGuard can exchange documents, reference data, workflow state, and draft records through APIs, webhooks, and adapters. The current portfolio build includes a sandbox accounting adapter only; the named external platforms below are integration examples, not deployed connectors.
QuickBooks, Xero, NetSuite, Sage, or SAP
Prepare draft bills, synchronize approved reference data, and return processing status through a controlled adapter.
Coupa, SAP Ariba, or internal purchasing systems
Read purchase orders, receipts, tolerance policies, supplier records, and approval ownership.
Email, Google Drive, SharePoint, S3, or secure upload
Receive invoices and supporting documents through APIs, webhooks, monitored inboxes, or storage events.
Slack, Microsoft Teams, email, or internal workflow tools
Notify the correct reviewer, link to evidence, and return human decisions without granting the model authority.
Supplier master, PO, receipt, identity, and access-control systems
Compare extracted claims with authoritative records and enforce organization-specific permissions.
Safety boundary
Connected systems may supply evidence and receive prepared draft records. A named human remains responsible for approval, and no integration grants the AI permission to execute payment.
Run five controlled scenarios and inspect every field, rule, and decision.
Built by Ariel
Ariel Magalso
AI Engineer
Philippines · Remote
Continue into the live workbench, inspect the source and evaluations, or contact Ariel about AI automation and workflow-engineering opportunities.
Open to opportunities
I build measurable AI-assisted workflows with deterministic safeguards, visible evaluation, and human review where the risk demands it.