Kanban StudiosKanban Studios
SELECTED WORK · 02

Mizan - the balance.

An autonomous AI case officer for housing-loan arrears rescheduling, built for the UAE Ministry of Energy & Infrastructure and the Sheikh Zayed Housing Programme. It weighs each beneficiary's hardship against policy and affordability - instantly, consistently, and with a full audit trail.

Mizan landing screen - An autonomous case officer
~5 days → instantManual review time, eliminated
100%Decisions backed by a full audit trail
EN / ARBilingual rationale on every case
THE PROBLEM

When a Sheikh Zayed Housing Programme beneficiary falls into arrears and requests a rescheduling of their housing loan, the case moves through a manual, roughly five-working-day review: an officer pulls the record, checks income and documents, validates against policy, drafts repayment options, and writes a recommendation.

It is slow, inconsistent between officers, hard to audit, and does not scale. Mizan turns that review into a near-instant, explainable, rule-compliant recommendation - escalating only the exceptional cases to a human.

WHAT IT DOES

A governed officer, not a chatbot.

01

Deterministic policy engine

The hard rules - income-deduction cap, repayment-period cap, document completeness, hardship evidence, active-request conflicts - are enforced in code, never by free-form LLM reasoning. A recommendation can never be hallucinated.

02

Candidate-plan solver

Generates concrete repayment options, drops anything that violates a rule, and ranks the survivors by sustainability and citizen burden - so the best compliant plan surfaces first.

03

Document intelligence

Classifies and reads uploaded documents, extracts income and other fields, and flags suspicious or missing paperwork - always returning structured, validated output.

04

Confidence & escalation

A single confidence score decides straight-through versus human review. Anything ambiguous, suspicious, or hardship-sensitive is escalated to an officer with a written reason and the supporting evidence.

05

Explainable, bilingual memos

Every case produces a written rationale in English and Arabic, with the exact policy rules and evidence each decision relied on.

06

Governed AI, human in the loop

LLMs only read documents and write prose. The decision logic is deterministic and auditable, and exceptional cases always land on a human officer's desk.

HOW IT RUNS

One governed pipeline.

Submitting a request runs the whole pipeline over a single, strongly-typed case state, streaming progress live:

Submit & assess - the governed pipeline runs node by node, streamed to the officer in real time.
Submit & assess - the governed pipeline runs node by node, streamed to the officer in real time.
IntakeRetrieve loan record
Doc auditRead the documents
Fraud checkFraud & duplicate
AffordabilityIncome analysis
Risk forecastRe-default risk
Policy solverCompliant plans
Human gateHuman-review gate
RationaleBilingual, finalize
The recommendation - a ranked, compliant plan with a confidence score, re-default risk, and candidate alternatives.
The recommendation - a ranked, compliant plan with a confidence score, re-default risk, and candidate alternatives.
THE OUTCOME

Explainable, every time.

Each case ends in a clear recommendation: the chosen repayment plan, a confidence score, a re-default risk read, and the ranked alternatives the solver considered. High-confidence, clean cases go straight through; anything sensitive is referred to an officer with the evidence attached - and every decision carries the policy rules it relied on.

BUILT WITH
FASTAPILANGGRAPHREACT + VITETYPESCRIPTPYDANTICPOLICY ENGINESQLITEANTHROPIC / GROQ
Chat on WhatsAppWhatsApp