WK
Wanjiru Kamau
Smallholder farmer · Kirinyaga · thin-file
Ethical multi-agent credit · financial inclusion

From risky to bankable — fairly.

AgriLift AI is an ethical multi-agent ecosystem that scores smallholder farmers using the data traditional banks ignore: mobile money patterns, produce sales, cooperative records, harvest trends, community trust and more. It expands access to credit while keeping lending responsible — with bias audits and human oversight built into every decision.

✓ 8 alternative indicators ✓ No collateral required ✓ Bias-audited (TRACK) ✓ Human-in-the-loop
<1 in 5
Kenyan smallholders with formal credit
the gap AgriLift closes
~70%
of Kenya's food from smallholders
yet under-served
8
Alternative credit indicators
no bank statement needed
5
Coordinated AI agents + 1 human gate
PRIDE-Loop oversight

The problem AgriLift solves

🚫 Invisible to banks

Smallholders have no collateral, no payslip, no formal credit history — so banks turn them down, even when they reliably repay.

⚠️ Or exploited by apps

Unregulated loan apps charge 30%+ monthly with repayment schedules that ignore harvest cycles — pushing farmers into debt traps.

✅ AgriLift's answer

Score the real signals of reliability, lend amounts and schedules matched to the harvest, and keep a human in the loop on every decision.

The five-agent ecosystem at a glance

🌾
Harvester

Gathers the 8 alternative indicators, with consent

🧮
Assessor

Scores creditworthiness from the indicators

⚖️
Auditor

Runs the TRACK bias & fairness check

🤝
Steward

Designs a harvest-matched, affordable offer

📚
Keeper

Logs the decision & reasoning (CBK-ready)

👤
Loan Officer

Human PRIDE-Loop — confirms every decision

Live tool · the 8-indicator scoring engine

Credit Score Builder

Drag each slider to reflect a farmer's real activity. AgriLift's Assessor agent computes a creditworthiness score in real time from eight weighted indicators — then the Steward agent translates it into a responsible, harvest-matched loan offer. No bank statement, no collateral.

⚠️ Demonstration scoring with illustrative weights. A production model is trained and validated on real repayment data and continuously bias-audited before any lending decision.

0
AgriScore /100

🤝 Responsible loan offer

Adjust the indicators to generate a harvest-matched offer.

⚖️ Fairness note

The Auditor agent checks that no single indicator dominates and that thin-file or female farmers aren't penalised by proxy.

Alternative agricultural credit indicators

The 8 Indicators

Traditional credit looks backward at a bank history most smallholders don't have. AgriLift looks at what farmers actually do — eight signals that genuinely predict whether a farmer can and will repay, each one consent-based and individually auditable.

Live prototype · the multi-agent credit decision

Agent Ecosystem — Live Demo

A loan request flows through five coordinated agents and one human gate: Harvester gathers consented data → Assessor scores it → Auditor runs the bias check → Steward designs an affordable offer → the human Loan Officer confirms → Keeper logs everything. Each agent is bounded by RANK, hands off via HUNT, and is watched by GUARD — and no loan is ever approved or denied by AI alone.

🌾
Harvester
Gather
🧮
Assessor
Score
⚖️
Auditor
Bias check
🤝
Steward
Offer
👤
Officer
Human gate
📚
Keeper
Log

RANK — bounded autonomy

Role · Authority · Notification · Kill Switch. The Assessor scores and the Steward proposes, but neither can approve or deny — that authority is reserved for a human Loan Officer. A kill code pauses all automated scoring instantly.

GUARD + TRACK

Every decision is logged with its full reasoning and a counterfactual ("would a male farmer with identical data score the same?"). Unusual approval-rate drops for any group trigger an anomaly alert.

// AgriLift AI agent ecosystem ready. Press Run.
Responsible lending · inclusion without harm

Responsible Lending

Expanding credit access is only good if it doesn't trap farmers in debt. AgriLift bakes responsibility into the product itself — not as an afterthought, but as a design constraint as important as approval rates.

🌱 Harvest-matched repayment

Repayment schedules align to the farmer's actual harvest and sales cycle — not arbitrary monthly dates. A farmer repays when income arrives, not before.

📉 Affordability ceiling

The Steward agent caps any offer at a level the farmer's cashflow can sustain — never the maximum the score allows. Over-lending is treated as a failure, not a sale.

🚫 No debt-trap pricing

Transparent, single, fair pricing — never the 30%+ monthly compounding of predatory apps. The full cost is shown in shillings before any farmer agrees.

🛟 Off-ramps & grace

If a harvest fails, the system offers restructuring and links to crop insurance — instead of immediate CRB listing. A bad season isn't a default.

📈 Credit that grows

Start small, repay, and your limit grows — building a real, portable credit history the farmer owns and can take to any lender.

🧑‍🌾 Always a human

No farmer is approved or denied by an algorithm alone. A loan officer reviews, can override, and must explain — and the farmer can appeal.

The responsible-lending test

Before any loan ships, AgriLift asks three questions: Can they repay it without harm? Does the schedule match their reality? Would we be comfortable explaining this loan to the farmer's family? If any answer is no, the offer is reshaped or declined — and the farmer is told why.

Lender view · portfolio health & inclusion impact

Lender Portfolio

What a SACCO or agri-lender sees: inclusion growing and risk staying controlled at the same time. Fairness and portfolio quality move together, not against each other.

+41%
Approvals for thin-file farmers
vs traditional model
3.8%
Portfolio-at-risk (PAR30)
held below 5% target
+58%
Women farmers approved
bias gap closing
2.1 days
Avg decision time
vs 3+ weeks manual

Inclusion vs risk — moving together

100
Traditional approvals
141
AgriLift approvals (+41%)
3.8
PAR30 risk % (controlled)

Recent decisions (sample)

FarmerScoreOfferDecision
Ethics & bias charter · responsible by design

Ethics & Bias Charter

A credit algorithm decides who gets economic opportunity. Done carelessly, it encodes old exclusions behind a "neutral" face. AgriLift's governance — built on the Safari curriculum frameworks — makes fairness a hard constraint, data sovereign, and every decision human-accountable.

ETHOS — Empathy · Transparency · Human Impact · Ownership · Sovereignty. Who is harmed if the model wrongly denies a reliable farmer; can the farmer see why; what a denial means for a planting season; the farmer owns their data and resulting credit history, not the platform.
TRACK — Training data · Representation · Amplification · Counterfactuals · Kill Switch. The core fairness engine: audits whether the model trains on data that under-represents women or remote farmers; runs counterfactuals ("would this score hold if the applicant were female / from a distant village?"); kills any scoring that shows disparate impact above threshold.
OASIS — Opt-in · Anonymization · Sovereignty · Intentional retention · Security. Every indicator is gathered only with the farmer's explicit consent in local language; mobile-money and sales data stays in-country under the Kenya Data Protection Act 2019; encrypted; deletable; never sold. DPA 2019 · CBK digital-lending rules
PRIDE Loop — Pause · Review · Interpretability · Disagreement rights · Elders council. No approval or denial is automated — a human loan officer confirms each one, with an explanation a farmer understands ("approved based on 3 years of consistent produce sales"). Farmers can dispute. Governance includes farmers, women's groups and cooperative reps.
HORIZON Scan — Historical harm · Opportunity cost · Ripple effects · Intergenerational · Zero-sum · Open futures · Non-human. Does cheaper credit lift households, or fuel over-borrowing? Does it strengthen cooperatives and seasonal resilience, or create dependency? Long view: build a credit history farmers own and carry forward.
RANK · TRAIL · HUNT · GUARD · CYCLE — the agent layer: bounded autonomy (no agent can approve a loan), consent-respecting memory, coordinated handoffs, GUARD safety rails (red-team test: "does the model penalise farmers who only deal in cash, or who are new to mobile money?"), and weekly self-improvement gated by a human officer's sign-off.

Fairness & responsible-lending checklist

ControlFrameworkStatus
AI scores & recommends but never auto-decides a loanRANK · PRIDECore rule
Per-decision counterfactual (gender / location)TRACKReady
Consent in local language before data is usedOASIS · ETHOSReady
Harvest-matched, affordability-capped offersResponsible lendingCore rule
No indicator may dominate the score (cap per signal)TRACK · GUARDReady
Disparate-impact monitoring & kill switchTRACK · GUARDPilot
AgriLift AI is a decision-support and inclusion tool — the lending decision always rests with a regulated human lender, and a score is not a guarantee of a loan. Scoring weights shown are illustrative; a production model must be trained on real repayment data, validated, CBK-compliant and continuously bias-audited before any live lending.
How it works · from data to dignity

How AgriLift Works

The full journey of a loan request — from a farmer with no bank history to an approved, affordable, harvest-matched loan — and how each safeguard fits in.

1. The farmer consents and connects data

With explicit, local-language consent, the Harvester agent gathers the eight alternative indicators — from M-Pesa records, cooperative logs, market attendance, supplier references and more. A farmer with only some of these still gets scored; the model never punishes a missing signal as if it were a bad one.

2. The Assessor scores creditworthiness

Each indicator is weighted and combined into an AgriScore. Crucially, no single signal can dominate — a thin mobile-money record can be offset by strong cooperative participation and community trust references, reflecting how rural reliability actually works.

3. The Auditor checks for bias

Before any offer, the Auditor runs the TRACK counterfactual: would an identical profile score the same if the farmer were female, younger, or from a remote area? If a gap appears, the decision pauses for human review rather than quietly excluding someone.

4. The Steward designs a responsible offer

Approval isn't "maximum loan." The Steward sizes the loan to the farmer's real cashflow, matches repayment to the harvest calendar, prices it transparently, and builds in grace for a failed season.

5. A human decides — and explains

The loan officer reviews the recommendation, confirms or overrides it, and the Keeper logs the decision with its reasoning for the farmer, the lender and the regulator. The farmer can appeal. Repaying grows a credit history they own and can carry anywhere.

This mirrors real Kenyan models — Apollo Agriculture, the Pathways/Fortune SACCO Farmers Credit Scoring Model, and FSD Kenya's cashflow-lending work — combined into one ethically-governed, multi-agent design.
AgriLift AI · Ethical agricultural credit ecosystem · Prototype — illustrative scoring, not a live lending product.