Skip to main content

Product features

A detection engine that shows its work.

LeakHunt is built around explainable findings: what was compared, why it was flagged, how confident the engine is, and what is still unknown.

Core detection

Three questions worth asking first.

Each detection type has its own evidence requirements and its own limits. LeakHunt keeps those boundaries visible.

Duplicate invoices

Exact matches use supplier, document number, currency, and total. Probable matches use supplier, amount, and currency within three days, with different document numbers.

  • Exact and probable match classes
  • Separate duplicate payment checks when payment evidence exists
  • Repeated imports do not create repeated findings

A duplicate invoice record does not prove the supplier was paid twice. Payment evidence is required for that claim.

Supplier price changes

The same supplier, item description, and currency are compared across confirmed price rows.

  • Previous price, current price, and absolute change
  • Percentage change with comparison dates
  • A minimum 5% and 0.01 increase before a finding is saved

A price increase is not automatically a financial loss. Quantity data determines whether an impact can be estimated.

Unusual expense amounts

Confirmed invoice, receipt, and expense amounts are compared with the same supplier's history in the same currency. A finding requires at least eight rows, an amount above the upper fence (third quartile + 1.5 × interquartile range), and at least twice the supplier's median.

  • The rule, supplier median, quartiles, and flagged row are shown
  • A minimum of eight confirmed rows prevents noisy flags
  • No savings estimate is made because the correct amount is unknown

An unusual expense is not proof of fraud or an incorrect payment. It is a reason to look closer.

Savings tracking

Three measures. Never blended into one.

Potential, confirmed, and realized savings answer different questions. Mixing them makes numbers look better and decisions worse.

Potential savings identified

The estimated impact of open findings, calculated from stored evidence and visible assumptions.

Stage 1 of the savings lifecycle

Savings confirmed by you

Only findings an authorized user has reviewed and confirmed. Software never confirms a finding.

Stage 2 of the savings lifecycle

Savings realized and reported

The amount your team reports as actually recovered or avoided, kept separate from estimates.

Stage 3 of the savings lifecycle

Every estimate shows its formula, input amounts, currency, time period, assumptions, and missing information. Overlapping findings are handled to avoid double-counting.

Available today: potential-savings estimates from open exact duplicate invoices, plus a recorded savings ledger where a reviewer writes down the exposure they accept as real and, later, what was recovered. Recorded amounts are entered by people, never inferred by detection.

Optional analyst

Ask questions. Keep the evidence.

The grounded analyst explains stored findings in plain language. It cites the source rows, refuses to invent amounts, and never confirms fraud or savings.

Local grounded mode

No external AI provider is required. The analyst explains findings, estimates, evidence, and imports with deterministic code on your server. Detection and every finding work without it.

Optional external provider

Administrators can connect an OpenAI-compatible provider on the server. Only the selected finding, counts, estimates, and up to 10 open findings are sent. Unsupported amounts and claims are rejected, with automatic fallback to local mode.

Either mode shows stored source rows with each answer. The analyst cannot change review status, mark a finding confirmed, determine fraud, or record savings.

Evidence and audit

Every finding can be traced back.

Findings are derived records, kept separate from imported source data. Opening a finding shows the records that produced it and the review decision your team recorded.

  • Finding ID, organization, type, and severity
  • Supplier and related invoice or transaction IDs
  • Detection method and confidence description
  • Estimated impact, formula, currency, and assumptions
  • Status, creation time, and update time
  • The review decision, its date, and any recorded savings

What LeakHunt will not do

  • Invent transactions, suppliers, or totals
  • Mark a finding as confirmed without a user
  • Call a price increase a financial loss by itself
  • Claim fraud or misconduct without evidence
  • Treat an uploaded document as instructions to follow
  • Expose one organization's records to another
  • Let one teammate quietly delete another's recorded saving

Organization isolation, server-side authorization, and Row Level Security are active for accounts, imports, findings, and recorded savings. Workspaces are shared with your team under owner, administrator, and member roles that the database enforces, not the interface. Original uploaded files are not retained; only validated rows, import metadata, and derived findings are stored.

See the dashboard your findings will live in.

CSV and XLSX imports, deterministic detection, finding review, CSV reports, and the grounded analyst are available. PDF and image OCR arrive later.

Sign in to the dashboard