Ariel Magalso · AI Engineer · Philippines · Remote
LedgerGuard is a solo-built accounts-payable workflow where AI reads variable invoices, deterministic code enforces financial controls, and people retain approval authority.
Solo
End-to-end ownership
Product design through deployment
5/5
Controls visible
Every decision stays inspectable
0.0%
Critical false clearances
Unsafe invoices cleared incorrectly, held-out set
$0
Payment authority
Public workflow cannot execute payment
Recruiter snapshot
Problem
Manual invoice preparation is repetitive, slow, and financially sensitive.
Solution
Evidence-linked AI extraction wrapped in deterministic controls and human approval.
My role
Solo product design, full-stack engineering, evaluation, operations, and deployment.
Stack
FastAPI, PostgreSQL, Claude, Docker, and a QuickBooks-style sandbox adapter.
Safety boundary
The workflow prepares work for review; it never authorizes payment.
Inspectable proof
Live demo, evaluations, architecture, operations, and source code are public.
Business case
Invoice operations · controlled automation · human authority
Built for organizations managing meaningful invoice volume across suppliers, locations, purchase orders, and approval routes—where speed matters, but accountable human judgment still matters more.
Operational risk
Accounts-payable teams repeatedly enter invoice data, reconcile purchase orders and receipts, check duplicates and supplier details, and route exceptions. The work is repetitive, but one missed discrepancy can delay operations or create financial loss.
Manual work · fragmented evidence · financial exposure
Controlled response
LedgerGuard uses AI to read variable invoice documents, then hands every financial decision to deterministic arithmetic, evidence matching, policy controls, and human approval.
AI extracts · code verifies · people approve
Operational value
Automation reduces manual preparation, surfaces exceptions sooner, preserves an inspectable audit trail, and gives finance teams more time for supplier communication and judgment-heavy review.
Faster preparation · clearer exceptions · visible audit trail
Where it fits
One control pattern, adapted to each company’s documents, policies, systems, and approval structure.
Reconcile recurring supplier invoices across buildings, cost centers, and property approvers.
Compare subcontractor and materials invoices with purchase orders, receipts, and project controls.
Standardize invoice intake and exception routing across many locations and local managers.
Match high-volume supplier charges against orders, deliveries, and approved commercial terms.
Prepare non-clinical supplier invoices for controlled review while retaining human authority.
Reduce repetitive invoice preparation for growing organizations with lean finance teams.
LedgerGuard can connect to accounting, procurement, document-intake, collaboration, and supplier-master systems. This portfolio build demonstrates a sandbox accounting adapter; external OAuth connectors are integration patterns, not deployed claims.
How the system stays safe
Featured scenario · LedgerGuard demo
The system extracts a Brightway invoice, aligns every field to evidence, recalculates the money, matches the purchase order and receipt, checks identity and duplicates, then prepares a draft for a human.
5/5
Deterministic controls visible
$0
Payment authority in the public workflow
Under the hood
Process
Read variable document layouts and propose structured fields.
Align evidence, recalculate money, match identities, and detect duplicates.
Assign a controlled outcome and the right reviewer using explicit policy.
Inspect evidence, correct fields, and approve or reject the prepared work.
About the builder
Ariel Magalso
AI Engineer
Philippines · Remote
I independently designed the workflow, data model, extraction schema, deterministic control engine, approval experience, evaluation harness, operational dashboard, and deployment.
Illustrative business case
Adjust the fictional inputs to understand the potential workflow effect. These values are intentionally separated from measured technical results.
AP hours potentially returned
Estimated net monthly savings (labor returned minus automation cost)
Demonstration assumptions, not customer outcomes. Defaults reproduce this project’s own stated fictional baseline (2,000 invoices/mo, 8 min manual prep, 60% straight-through → 160 AP hours/mo).
FAQ
Three systems, one argument.
Each system automates judgment work an LLM can confidently get wrong. As the consequence grows, deterministic controls and human authority take over.
Recoverable Irreversible
Open to opportunities
I build measurable AI-assisted workflows with deterministic safeguards, visible evaluation, and human review where the risk demands it.