VasooliRevenue recovery control plane
Verifying audit chain
000
sha256 · 0000000000000000000000000000000000000000

Every decision the agent makes is hash-chained. Nothing here is a screenshot.

Razorpay AI Buildathon 2026 — Track 3

Anyone can reportthe money that came back.We report the moneythat wouldn’t have.

Vasooli holds back a random fifth of every at-risk case and never touches it. The distance between that cohort and the one it works on — with a confidence interval, on a hash-chained ledger you can re-verify yourself — is the only number it takes credit for.

At risk
Policy gate
Recovered
Treatment · 80%Holdout · 20%
The problem with gross recovery

A customer whose card was declined by a downed issuer will often retry on their own an hour later. If your agent emailed them in between, it takes the credit. Gross recovery cannot tell those two cases apart, which means it cannot be checked, reproduced, or argued with.

What everyone reports
₹4.2L
Gross recovered. Every rupee that came back after the agent touched a case, whether or not the agent had anything to do with it. Unfalsifiable by construction.
What Vasooli reports
₹2.9L
Incremental, at 95% confidence, measured against a cohort we deliberately left alone. The rest of the gross figure is money that was always coming home.

The difference is not a rounding error. It is the difference between a number you can put in a board deck and a number you can defend in one.

The path from signal to action

This ordering is the safety story. A language model proposes a diagnosis and writes copy; a bandit chooses the intervention; deterministic code decides whether anything happens at all. Nothing skips the gate.

  1. 01

    Detect

    CUSUM change-point on per-cohort success rates, checkout TTL expiry, mandate grace periods, invoice aging buckets. Real Razorpay error codes throughout.

    Deterministic
  2. 02

    Assign

    sha256(case id) mod 100 decides treatment or holdout. Stable across retries and restarts, and re-checkable by anyone holding the case id.

    Deterministic
  3. 03

    Diagnose

    The model proposes a root cause, a confidence, an evidence code and a segment. Schema-validated on arrival. It cannot pick an action and cannot move money.

    Model
  4. 04

    Plan

    Thompson sampling over Beta posteriors picks the playbook arm. The model writes the copy for the arm that was already chosen — it never chooses.

    Bandit
  5. 05

    Gate

    Every rule runs, every verdict is kept, most restrictive wins. Nothing reaches the executor without passing here.

    Deterministic
  6. 06

    Approve

    Discounts, fee waivers and anything above the auto-approve threshold park in an inbox until a person decides. Rejection stops the case dead.

    Human
  7. 07

    Execute

    Durable steps with per-step retry and reverse-order compensation. A failure half way through unwinds what already happened.

    Deterministic
  8. 08

    Measure

    Wilson and Newcombe intervals, mSPRT for an always-valid p-value. Outcomes feed straight back into the bandit's posteriors.

    Deterministic
The gate

Every rule runs on every case. Each returns a verdict and a reason, all of them are written to the ledger, and the most restrictive one governs — a BLOCK is never quietly downgraded because something else returned PASS.

Hard stop on terminal cases
hard_stop_terminal
Case already recovered, failed or stopped
BLOCK
Economic viability
economic_viability
Action costs more than the amount it could recover
BLOCK
Touch cap
max_touches_per_case
Three outreach attempts already made
BLOCK
TRAI quiet hours
trai_quiet_hours
Local time between 21:00 and 09:00
DEFER
Cool-off window
cool_off_between_touches
Under four hours since the last touch
DEFER
Human sign-off
human_approval_high_risk
Discount, fee waiver, or above the auto-approve threshold
NEEDS_APPROVAL

Twenty-six tests cover this file alone, one per rule per verdict, plus the combinations where two rules disagree.

The number

Connecting
₹0
95% CI ±₹0
p = 0.000

Recovered because the agent acted, not recovered while it happened to be watching. The p-value is always-valid, so this page can be refreshed mid-pitch without invalidating the statistics.

Treatment
n=92
30.4%
₹1.18L+20.9 pts incrementalcollecting
Holdout
n=21
9.5%
Gross recovered
₹1,67,690
Cases detected
113
Cases recovered
30
Stopped by policy
0
The audit trail

Every transition, verdict, diagnosis and step outcome is appended to a hash-chained ledger. The audit page re-hashes the entire chain in your browser with Web Crypto and tells you what it found — it does not ask the API whether the API is trustworthy.

Chain
SHA-256
Serialisation
Canonical JSON
Tamper test
In the suite
Case timelinechain valid
  1. detectedorchestrator
    prev · eb3da79e2cd9
  2. diagnosedllm
    prev · 7f41c76c91c3
  3. plannedbandit
    prev · b4639a53c1a2
  4. policy_evaluatedpolicy
    prev · c885bf584d72
  5. approvedhuman
    prev · 4c29bec89a4c
  6. execution_resultexecutor
    prev · 9d3bb6a1240b
  7. recoveredorchestrator
    prev · 80876955fe75
What it is made of

Tests
160
across 13 packages, none touching the network
Statistics
Hand-rolled
Wilson, Newcombe, CUSUM, Thompson, mSPRT
Money
bigint paise
no float arithmetic anywhere in the system
Runs offline
By default
mock model and payment fake, seeded and deterministic
Replay
Byte-identical
same seed reproduces the same batch exactly
Database
None
in-memory engine state; nothing to stand up first