Verification engine

Every incoming message runs through the same eight stages, in order, before a payment can be approved.

The pipeline

One message in, one decision out. Regex runs first because it is deterministic and free; the LLM runs in parallel and the two results are only trusted when they agree.

  1. 01

    Incoming message

    An email, SMS, webhook or API event arrives from a connected verification source.

  2. 02

    Preprocessing

    Strips signatures and quoted replies, normalises whitespace, encoding and currency symbols.

  3. 03

    Regex parser

    Matches the message against the active parse rules for that sender, field by field.

  4. 04

    LLM parser

    Sends the message to the configured model with the JSON extraction schema.

  5. 05

    Result comparison

    Diffs the two parses field by field — amount, currency, transaction ID, receiving account.

  6. 06

    Order matching

    Searches pending orders inside the payment window for an exact amount and reference.

  7. 07

    Risk / confidence calculation

    Weights parser agreement, the matched custom rules and the time window into one score.

  8. 08

    Decision

    Auto-approve, route to manual review or reject — then notify the merchant over webhook.

Worked example — order #ORD-10428

The 16:04 bank notification for John Doe’s ₹2,499 UPI payment, parsed by both engines in 1.4 seconds.

Incoming message

HDFC Alerts Gmail · 1 Sep · 16:04:04
INR 2499.00 credited to account XX4921. UPI Ref 928473928.

Regex result

Rule · HDFC UPI alerts
Amount
2499.00
Currency
INR
Transaction ID
928473928
Account
XX4921
Confidence
100%

AI result

gpt-4o-mini · 1.2 s
Amount
2499
Currency
INR
Transaction ID
928473928
Account
XX4921
Transaction type
Credit
Confidence
99%

Comparison — match

Regex and AI agree on amount, currency, transaction ID and receiving account. Combined confidence 99%.

Auto-approved ≥ 95%