Veriflow: B2B Revenue Leakage & Invoice Auditing System

Project Overview

Veriflow is a production-oriented dual-model machine learning compliance platform that automatically monitors B2B transactions, identifies pricing discrepancies, and flags suspicious shipping rates to eliminate revenue leakage in enterprise supply chains. Connected to a 424MB relational SQLite database containing millions of retail wine and spirits distribution logs, the system helps corporate financial auditors isolate overcharges before payments are finalized.

Situation

Large retail and logistics companies process thousands of B2B transactions every month. Due to highly complex contracts, multi-tiered shipping rates, and manual processing errors, enterprise organizations routinely overpay suppliers by 1% to 5% annually. For Bibitor, LLC (the distribution catalog represented in the relational schema), auditing millions in billed volume manually is extremely resource-intensive, making automated leakage detection crucial.

Task

The goal was to engineer a production-ready machine learning framework and an interactive compliance dashboard that connects directly to the historical transaction database, extracts complex relational logs, builds highly expressive features, and implements a robust dual-model architecture to forecast expected logistics costs and classify high-risk billing anomalies.

Actions

• Database Integration & Join Querying: Connected the platform to the SQLite data store, performing complex left-joins across purchases, vendor_invoice, purchase_prices, and inventory tables.
• Expressive Feature Engineering: Derived features to isolate overcharges, including:
  • Price Variance: Discrepancies between contract purchase orders and billed invoice totals.
  • Quantity Discrepancies: Tracking if quantity billed exceeded quantity ordered.
  • Freight-to-Invoice Ratio & Freight-per-Unit: Normalizing shipping costs to detect padded shipping bills.
  • Billing Delays: Calculating days elapsed between purchase orders and invoice dates.
• Dual-Model Machine Learning Pipeline:
  • Expected Cost Regressor: Trained a Random Forest Regressor ($R^2 = 0.9661$) to predict fair, baseline freight charges.
  • Risk Classifier: Trained a Random Forest Classifier (90% weighted F1-Score) to flag compliance violations.
• Edge-Case Safeguarding: Implemented robust exception handling including unknown vendor ignoring via one-hot encoding setups, zero-division offset protection (+1e-5), and PO fallback layers.
• Compliance Dashboard: Developed a premium Streamlit interface equipped with executive variance stats, an interactive simulation auditing desk, and a filtered SQL ledger with CSV exporting.

Results

Veriflow analyzed 5,543 enterprise transactions totaling $21,080,266.39 in audited billing volume. The dual-model compliance pipeline successfully isolated 86 extreme freight overcharge anomalies (where shipping bills exceeded >0.8% of invoice values), offering immediate audit clarity and demonstrating significant potential for corporate savings.

Tech Stack

• Streamlit: Framework for the premium compliance dashboard UI.
• Scikit-learn: Employed for Random Forest models, preprocessors, and training validation metrics.
• SQLite: Relational relational database backend.
• Pandas & NumPy: Libraries for processing relational datasets and complex array metrics.
• Python: Scripting core for data extraction and modular inference APIs.