LedgerGuard AI

## Inspiration

In corporate finance, accounts payable (AP) and treasury operations are plagued by friction and financial leakage. Over 70% of finance analyst hours are spent manually cross-referencing supplier
invoice line items against purchase orders (POs) and warehouse receiving slips (GRNs). Even worse, organizations lose billions annually to uninspected short-shipments, subtle supplier price creep
exceeding contracts, and catastrophic Business Email Compromise (BEC) wire fraud where attackers hijack vendor remittance routing numbers.

We wanted to build an autonomous agent system for the Office of the CFO that doesn't just digitize invoices, but actually acts as a forensic controller—intelligently evaluating variance tolerances,
catching fraud before funds leave the bank, and giving the CFO an intuitive human-in-the-loop review cockpit. ────── ## What it does

LedgerGuard AI is an autonomous multi-agent financial controller and fraud sentinel for accounts payable:

Automated 3-Way Reconciliation: Automatically correlates PO line items, warehouse Goods Received Notes (GRN), and invoices down to quantities and unit prices within customizable vendor tolerance
thresholds.
Fraud & Remittance Sentinel: Cross-checks bank routing numbers and account details against verified corporate Vendor Master records, instantly halting payments on unverified account switch attempts.
Touchless ERP Posting: Automatically approves flawless matches and generates SAP / NetSuite-ready double-entry ledger postings without human touch.
Autonomous Supplier Dispute Generation: When variances or short-shipments occur, the agent drafts formal, polite dispute notices to supplier AR demanding credit memos or corrected billings.
Human-in-the-Loop (HITL) Review Cockpit: Gives controllers a 1-click decision center to approve overrides, dispatch disputes, or reject fraudulent invoices, complete with mandatory audit justification notes.
Immutable Audit Trail: Keeps a timestamped log of all agentic and human decisions for SOX and GAAP compliance. ────── ## How we built it

• Core Orchestration & Development: Built end-to-end using AO (Agentic Orchestrator) from initial architecture planning, code authoring, dependency troubleshooting, to git version control. • LLM Reasoning: Integrated TensorMux's GLM-4.7-Flash (30B MoE agentic model) via its OpenAI-compatible endpoint (https://api.tensormux.com/v1) to synthesize complex forensic reviews, determine risk
severity, and craft natural supplier dispute correspondence. • Backend Architecture: High-speed FastAPI (Python 3.12) backend using Pydantic v2 models for strict financial validation, vendor master tolerance scoring, and automated double-entry journal generation. • Frontend Dashboard: A responsive dark-mode CFO Command Center with real-time KPI metrics, line-item 3-way variance diff tables, dispute previews, and interactive review modals. ────── ## Challenges we ran into

• Balancing Strict Determinism with Agentic Reasoning: Financial accounting cannot tolerate hallucinations or fuzzy math. We designed a dual-layer architecture where strict deterministic rules handle line-item tolerance math and bank hash checks, while TensorMux GLM-4.7-Flash handles qualitative risk synthesis, executive recommendations, and supplier dispute drafting. • Simulating Realistic Enterprise Data: Crafting realistic enterprise edge cases (partial shipments, price creeps, and subtle banking routing switches) that mirror real-world corporate fraud and
operational failures. • Cross-Platform Tooling & Fast Scaffolding: Managing native Windows package bindings during rapid full-stack scaffolding in AO, pivoting quickly to zero-friction embedded architectures to keep
development fast and reliable. ────── ## Accomplishments that we're proud of

• True End-to-End Autonomous Workflow: Flawless invoices are matched, reconciled, and posted to ERP ledgers in sub-seconds with 0 human intervention required. • Measurable Financial ROI: In our test suite, LedgerGuard evaluated over $118,000 in spend, auto-approved standard orders, and successfully blocked over $58,000 in short-shipment overbillings and
fraudulent remittance redirects. • Seamless Human-in-the-Loop Experience: Designed an intuitive cockpit where finance leaders retain executive control through 1-click approvals and SOX-compliant audit logging. • 100% AO Built: Successfully planned, built, tested, and shipped the entire project natively within AO. ────── ## What we learned

• Real-world finance automation requires deep respect for existing accounting workflows: accountants don't want a "black box" that silently pays invoices; they need clear audit trails, variance
transparency, and actionable dispute drafts. • GLM-4.7-Flash via TensorMux is fast and capable at agentic reasoning, structured JSON outputs, and domain-specific financial tone. ────── ## What's next for LedgerGuard AI

• Direct ERP Connectors: Native OAuth sync for NetSuite SuiteTalk, Workday, and QuickBooks Online APIs. • Multi-Modal Document OCR: Direct ingestion of scanned PDF receipts, delivery manifests, and multi-page supplier bills via multi-modal agent vision. • Automated Supplier Email Loop: Integrating direct bidirectional email/Slack dispatch so suppliers can reply with credit memo PDFs that the agent re-evaluates automatically.