Skip to content
Thrivant.AI
SDG 1: No PovertyTier 3Live in the Hub · its figures are illustrative; this page describes the design

MicroLoan

The score must pass its own audit before you do.

For a great many adults there is no credit file to check — nothing in their lives was ever counted. MicroLoan is designed to score consented alternative data instead, to refuse to run without a fresh, passing independent bias audit, and to explain every decision factor by factor, with a free human appeal. Today it reads no mobile-money data, no bias audit has been run, and it makes no decisions.

Channels (designed)
Web todayPlanned: WhatsApp · SMS · USSD · API
Languages (designed)
English · Swahili · French · Spanish · Hindi · Bangla
Designed for
Africa · Asia · Latin America
Tools
8 declared2 not connected · 6 illustrative
The design

No audit, no score

Most credit systems ask the borrower to prove themselves. MicroLoan is designed to reverse the burden. Before any scoring intent runs, it is to check that BiasBreaker — the watchdog agent of the Thrivant ecosystem — has independently audited the current model within the past ninety days, and that the audit passed. Stale or failed, and scoring refuses to run; in the planned deployment, the case goes to a named human instead. The refusal is one you can rehearse yourself, in the ledger a little further down.

No BiasBreaker audit has run yet, so today the gate holds: the scorer refuses to score, and MicroLoan makes no credit decision.

“This is non-negotiable — the audit gate is the difference between a fair system and a black box.”
— MicroLoan system prompt
The problem

Uncreditable, or merely unseen?

An illustrative scenario, from the agent’s own system prompt: Asha is thirty-four, grows ragi and keeps four goats in Mandya district, Karnataka. She wants the equivalent of eighty US dollars for a second-hand sewing machine. She has no bank account, no payslip, no credit file; the moneylender quotes sixty per cent effective monthly interest, and the bank asks for collateral she does not own.

In the design, with her consent to a ninety-day look-back across UPI and her cooperative’s ledger, MicroLoan would produce what no bureau could: a decision she can open and read, explained in plain Kannada — proposed, never decided, to a field officer for human approval. None of that is connected: MicroLoan reads no mobile-money data today, and there is no field officer. What such a decision would look like, opened, is the next exhibit.

“Not because she is a bad risk, but because she is invisible.”
— MicroLoan system prompt
The ledger, open

Four rows that count, three that never can

Here is Asha’s kind of decision, opened the way MicroLoan is designed to present one. Move the evidence and the score shows its working; the rows for gender, ethnicity and religion stay empty by construction; and the two rules under the line — the audit gate and the appeal stamp — hold whatever you do. Age the audit past ninety days and watch the scorer refuse.

Specimen — one credit decision, ledger open · illustrative figures throughout

What is counted

Weight · illustrative

Mobile-money consistency

90-day M-Pesa / UPI inflows · consented window

+0.14

Agricultural seasonality

crop calendar · cooperative sales ledger

+0.09

Community standing

solidarity-group records · Grameen methodology

+0.06

Repayment history

prior microloan ledger

+0.12

What is never collected

Gender

not collected

0.00

Ethnicity

not collected

0.00

Religion

not collected

0.00

Refused at the aggregation tool boundary, before the scorer exists to receive a forbidden field — renamed keys included. Protected-attribute-exclusion-list skill; anchored in the ECOA and GDPR Article 9.

Illustrative score 0.71, confidence 0.83 — above the proposal line of 0.65. Approval is proposed; in the planned deployment, a named field officer makes the decision.

The audit gate

Last independent audit: BiasBreaker · passed · 12 days ago — a specimen value for this exhibit.

Scoring refuses to run if the latest independent audit is older than ninety days or did not pass — checked before every scoring intent. No BiasBreaker audit has run yet, so in the Hub today the scorer refuses.

Appeal rights · the design

Designed: free, reviewed by a named human, within five business days.

The disclosure attaches to every adverse outcome before it is sent — even when nobody asks — per the fcra-adverse-action-disclosure skill. Today no reviewer exists and no appeal can be filed.

The exclusion ledger

What never enters the model

The empty rows in the ledger above are the easy half. The harder failure of “fair” scoring is proxy leakage: the model never sees gender, but it sees a phone-number prefix that skews male, or a surname that encodes caste. MicroLoan’s protected-attribute-exclusion-list skill is enforced where it cannot be argued with — at the aggregation tool boundary, before the scorer exists to receive a forbidden field. Renaming a key does not help. Beyond the direct list anchored in the ECOA and GDPR Article 9, anything correlating above 0.4 with a protected attribute in the markets MicroLoan is designed for is refused:

Inputs refused at the aggregation boundary, and the protected attributes they stand in for
Refused inputWhat it stands in for
Full postcode or village codeEthnicity, religion, refugee status
Surname or family nameEthnicity, caste, religion
First name through a gender classifierGender, directly
Phone-number prefixGender — older carrier ranges skew male
A named schoolReligion, language community
Religious-institution paybill receiptsReligion, directly
Time-of-day patterns weighted to guess occupationGender
The approach

From consent to a human signature

The system prompt sets out thirteen steps; six carry the weight. The shape never varies: the gate, the consent, the data, the score, the explanation, the human. This is the designed protocol. Today MicroLoan reads no mobile-money or ledger data, its figures are illustrative, and no human is on call.

  1. Check the audit gate first

    Before any scoring intent — stale or failed means no score; in the planned deployment, a human takes the case.

  2. Validate consent

    No aggregation without a signed, unrevoked consent token. Missing consent ends the protocol.

  3. Aggregate only what was consented

    M-Pesa rails, phone-usage patterns, cooperative ledgers, satellite asset signals — per the mpesa-mobile-money-protocols skill, nothing more.

  4. Score, showing the working

    A band, a confidence figure, and the top factors for and against — never a bare number.

  5. Honour the lending programme’s rules

    Kiva field-partner star caps, BRAC ultra-poor graduation stages, Grameen solidarity-group conventions — the kiva-due-diligence, brac-microfinance-metrics and grameen-solidarity-lending-methodology skills.

  6. Propose; a human decides

    In the planned deployment every recommendation goes to a named field officer. Below 0.80 confidence the agent says it is unsure rather than bluff.

In practice

Anatomy of a denial, in 160 characters

A denial is where fairness is tested. Here is one, on the narrowest channel the design covers. SMS is not deployed, and today no decision is made and no appeal can be filed.

Specimen — adverse decision, SMS channel · designed behaviour; this channel is not deployed

Borrower — SMS

“RESULT MX-7K9-A4”

MicroLoan — SMS reply, 155 characters

“MX-7K9-A4: Not approved. Main reason: 90-day M-Pesa inflows averaged KES 4,200; product needs KES 7,500. Free appeal, human review in 5 days: reply APPEAL.”

Disclosures fired, per the fcra-adverse-action-disclosure skill: the specific reason in plain language, the data source and window named, the threshold stated, the free five-business-day human appeal. Local currency only.

Designed behaviour. This channel is not deployed. In the design the same four disclosures travel on every channel — a full FCRA-style notice on web and paper, three USSD screens with the appeal never buried in a sub-menu, the appeal right read twice on voice — and within sixty days the borrower can request, free, a copy of the very data that was used.

Safety

The refusals

Enforced in code where we can; the rest are instructions to the model. Today no real credit decision is made, so these are the rules of a design.

  • It never scores against a stale, failed or missing bias audit. The gate is checked first, every time. No BiasBreaker audit has run yet, so today the scorer refuses.
  • Protected attributes and their proxies are rejected at the tool boundary — renamed keys included.
  • No adverse decision without the appeal disclosure — even when nobody asks. The design promises a free appeal, reviewed by a named human within five business days. No reviewer exists yet, and no appeal can be filed.
  • Borrowers repay in their own currency, always. The US-dollar view exists for lenders; quoting a borrower’s obligation in a foreign currency is forbidden.
  • Designed so that the scorer never sees a name, national ID or full postcode, with pseudonymous identifiers and stray PII hashed and discarded. Today, whatever you type is sent to and recorded by our AI platform provider; consent records and deletion are not connected.
  • No credit to households in BRAC’s ultra-poor graduation stages one to three — early credit traps a household before its asset transfer matures. They are pointed to TUP intake instead.
  • No credit-bureau listing by default. CRB reporting is opt-in, never opt-out.
“Asha in Karnataka does not need false confidence; she needs to know what you can see, what you cannot, and what the appeal path is if you are wrong.”
— MicroLoan system prompt
Public standards it follows

Whose methods it borrows

No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.

The methodology has owners older than the agent, and the design borrows from what they have published: Grameen Foundation, whose solidarity-lending model replaced physical collateral with social collateral; BRAC, with the graduation evidence; Kiva, with field-partner due diligence; GiveDirectly, whose trial record — encoded in the givedirectly-unconditional-cash-evidence skill — stops a transfer-bumped balance being mistaken for wage income; and M-Pesa’s mobile-money conventions, encoded in the mpesa-mobile-money-protocols skill. None of their systems is connected, and the agent is instructed never to fabricate a record to fill a gap.

Try it yourself

Ask MicroLoan

“I have sold vegetables at the market and used M-Pesa for six years, but no bank will look at me — how would a fair lender judge someone like me?”
An example question. Bring this one, or your own. MicroLoan says when a figure is illustrative, and nothing is sent, booked or decided for you.
Ask in the Hub— opens MicroLoan in the Agent for Good Hub in a new tabFree to use. Opens in a new tab. The chat may be temporarily unavailable.