11-stage AI pipeline · Bank · GL · AP

Catch financial fraud
your auditors miss

Upload a CSV or Excel ledger. Auraium runs 25+ detection rules, statistical outlier scoring, and unsupervised ML — then delivers plain-English audit reports in under 60 seconds.

✓ No credit card required · ✓ 10,000 transactions free · ✓ Results in under 60 seconds
Duplicate invoices 📊Velocity spikes 🔍New counterparty fraud 🤖ML anomaly scoring 📒Journal entry manipulation 🧾Threshold splitting
11 Pipeline stages
25+ Detection rules
<60s Processing time
9 Ledger types supported

Why Auraium?

Rule-Based

Fast, deterministic checks for known suspicious patterns (duplicates, blacklisted accounts, threshold breaches).

📊

Statistical

Robust outlier scoring using distribution-aware math for volume, frequency and velocity deviations.

🤖

ML Inference

Models identify complex patterns and subtle fraud that rules miss — continuously improving with data.

Works with every ledger format

Auraium auto-detects your ledger type and adapts the entire pipeline — rules, statistical baselines, ML features, enrichment context, and LLM prompts — to match.

🏦

Bank Statements

Detect counterparty fraud, velocity spikes, pass-through structuring, round-amount patterns, and weekend anomalies.

New counterparty Velocity spike Pass-through Self-transfer
📒

General Ledger

Flag journal entry anomalies, period-end manipulation, unusual department-account combinations, and misclassification risk.

Period-end posting Misclassification Unusual GL account
🧾

Accounts Payable

Catch duplicate invoices, invoice splitting near approval thresholds, first-time vendor fraud, and vendor concentration risk.

Duplicate invoice Threshold splitting New vendor
📬

Accounts Receivable

Identify revenue recognition anomalies, fictitious customers, credit manipulation, and collection irregularities.

Fictitious customer Revenue timing Credit override
🧳

Expense Reports

Surface personal charges, duplicate reimbursements, round-amount padding, and out-of-policy spending patterns.

Duplicate claim Round-amount padding Policy breach
🔄

Cash Flow / Reconciliation

Flag timing gaps, unreconciled items, unusual sweep patterns, and discrepancies between bank and book balances.

Timing gap Unreconciled item Sweep anomaly
📦

Inventory & Assets

Detect phantom inventory, unusual write-offs, valuation irregularities, and misclassified capital vs. operating expenditure.

Phantom inventory Write-off spike Valuation shift
💼

Payroll Register

Catch ghost employees, duplicate pay runs, excessive overtime, and compensation deviating from peer benchmarks.

Ghost employee Duplicate pay Overtime spike

Auto-Detect

Not sure which type? Upload anything — Auraium classifies your ledger automatically and applies the right detection profile.

Zero config AI classification Any format

11-Stage Detection Pipeline

Three phases run in sequence. Within Phase 2, all three detection engines run in parallel — their scores fused by the Decision Engine.

Phase 1 — Ingest & Classify
📥
Ingest & Validate

Auto-detect CSV or Excel. Normalize columns, validate schema, map to standard fields.

🔍
Ledger Classification

Identify: Bank Statement, General Ledger, or Accounts Payable. All downstream stages adapt accordingly.

Phase 2 — Parallel Detection runs simultaneously
Rule Engine

25 rules: duplicates, velocity spikes, threshold clustering, blacklist matching, cross-field logic.

📊
Statistical Detector

4-method ensemble — rolling z-score (40%), personal baseline (40%), peer group (15%), velocity (5%).

🤖
ML Ensemble

Isolation Forest + PCA reconstruction. Unsupervised — no labeled training data required.

Phase 3 — Enrich, Decide & Explain
📈
Context Enrichment

90-day account baselines, counterparty history, new-party flag, pass-through & velocity enrichment.

⚖️
Decision Engine

Agreement amplifier: 3 detectors = 1.6× boost, 2 = 1.3×. Flag rate capped between 0.1%–5%.

🛡️
Fraud Validator

3-tier recheck: heuristics → small ML → LLM. Each flagged transaction is confirmed or cleared.

📝
LLM Explainer

8-field audit report: tldr, what happened, why unusual, red flags, data gaps, verdict, action, next steps.

Inside each detection engine

Scroll through the five core engines — each outputs a score, a reason, and contributes to the final decision.

Rule Engine

25 rules: duplicate IDs, velocity spikes, amount clustering near approval thresholds, blacklist matching, and cross-field conditional logic. Outputs: rule_score (0–100) + reason.

📊

Statistical Detector

4-method weighted ensemble: rolling z-score, personal baseline ratio, peer-group comparison, velocity spike — weighted 40 / 40 / 15 / 5. Flagged if composite > 1.0. Outputs: stat_score + reason.

🤖

ML Ensemble

Unsupervised: Isolation Forest + PCA reconstruction error. No labeled data required. Flags the top 5% of transactions by combined anomaly score. Outputs: ml_score + feature contributions.

⚖️

Decision Engine

Fuses three detector scores. Agreement amplifier: all 3 detectors triggered = 1.6× confidence boost, 2 detectors = 1.3×. Flag rate enforced within 0.1%–5% bounds. Outputs: FRAUD_HIGH / MEDIUM / LOW / AUDIT_ONLY.

📝

LLM Explainer

Profile-aware (bank / GL / AP). 8-field audit report per transaction: tldr, what happened, why unusual, red flags, data gaps, verdict (Fraud / Likely / Uncertain), recommended action, and 3 specific next steps.

What an audit report looks like

Every flagged transaction gets a structured 8-field report — written in plain English, not jargon.

ID: trxn_00012345 Amount: ₹1,20,000 Date: 2025-11-26 (Saturday) Counterparty: Apex Trading Co. (new)

"₹1,20,000 payment — 12× this account's typical transaction size — sent to a first-time counterparty on a weekend."

⚠ Likely Fraud 🔍 Investigate 87% confidence

Privacy by design

We minimize data shared with external services and always prefer on-premise or private inference for sensitive fields.

🔒

Minimal fields to API

Only column headers, metadata tags, and hashed IDs are sent to external services. Raw PII — names, account numbers — is never sent without being obfuscated, and only when explicitly allowed.

🖥️

Local-first inference

Rule and statistical engines run entirely on your infrastructure. If cloud ML is used, data is pseudonymized before leaving your environment, and no transaction data is retained after processing.

📋

Auditable trace

Every LLM call is logged with a prompt hash, model version, and minimal field bleed. You can reproduce any explanation, verify any decision, and export a full audit trail per run.

FAQ

What happens while mapping fails?

If automatic mapping fails, Auraium will (optionally) call an internal LLM helper with only column names — not values — to suggest mappings. You always see suggested mappings before they’re applied.

How do we measure precision / false positives?

Auraium tracks precision, recall, and false positive rate per rule and per model. You can view suite-level metrics on the dashboard and set thresholds to meet your tolerance.

Can I opt-out of cloud ML?

Yes — Auraium supports on-premise models and private inference. Cloud ML is optional and used only after explicit configuration.

Ready to try clearer, safer detection?

Get started Contact sales