Last updated By Sho

Field note / 16 November 2026 / real product / Bila Boleh

The calculator was never the hard part.

Bila Boleh began with an uncomfortable question: if a spreadsheet already calculates retirement, what is genuinely worth building?

Real productDated field note16 Nov 2026 · Sho
Bila Boleh / product-engineering field note
What you'll learn

Distinguish a calculator from a persistent, connected financial planning system and see where AI helps without becoming the arithmetic engine.

Bring / required

Nothing personal. The worked checks below use fictional figures; do not upload actual financial records.

What you'll do

Inspect the architecture, challenge one double-counting failure, and define evidence that would make the system trustworthy.

Take away

A reusable checklist for state, deterministic calculations, assumptions, privacy and verification in AI-assisted product builds.

Why another retirement calculator?

A good spreadsheet can already calculate compounded returns, inflation, drawdown and a retirement date. It can preserve a sophisticated model, too. A prettier dashboard on its own would not be a strong reason to build another product.

The actual question was different: can someone keep one coherent picture of the money they own, the debts they owe, the income they receive and the life they want — then return next month to ask a different question without rebuilding the plan?

That is the difference Bila Boleh is attempting. Not one answer, but a plan that can be revisited, corrected and challenged.

The problem worth solving

Don't build another calculator. Build a plan whose numbers still agree after the user changes something.

From isolated calculators to connected state

01 / INPUTS Money lives in pieces.

EPF · savings · holdings · income · debts · spending

02 / MODEL One saved plan.

Recorded facts + stated assumptions + deterministic rules

03 / QUESTIONS What can change?

Retirement age · spending gap · accessible capital · scenarios

Conceptual explanation, not a live financial chart or a claim that AI calculates retirement outcomes.

The public Bila Boleh planner brings EPF/KWSP, other savings, shares and dividends, household income, spending, assets and liabilities into a shared retirement outlook. The point is not to count how many calculators exist. It is to make the same underlying values drive the overview and the detail pages.

LayerResponsibilityWhat must not happen
Recorded factsUser-entered balances, assets, liabilities, income and spending.Present an entered number as a live verified bank balance.
Planning assumptionsInflation, returns, contribution rates, retirement ages and spending stages.Quietly turn a chosen assumption into a guarantee.
Calculation engineDeterministic projections, availability dates, cash flow and asset values.Count the same transfer, repayment or dividend twice.
Saved planKeep inputs and scenarios consistent across pages and later visits.Force someone to re-enter their working model every session.
PresentationExplain the outcome, the shortfall and the assumptions behind the chart.Confuse net worth with cash that can be spent today.

The hard parts appear between the rows. Dividends can be reinvested or spent. Loan principal reduces debt but is not new wealth. A property can add net worth without funding next month's expenses. Retirement funds may have access-age restrictions. These are product and financial-model decisions, not prompt-writing problems.

What AI should do — and what it should not

WORK WITH AIDEFINE

Frame the real job

Use AI to challenge the product brief, identify missing scenarios, compare alternatives and explain assumptions. The person remains responsible for choosing what is true and useful.

GIVE AI WORKBUILD

Implement bounded work

Let coding agents implement UI, parsers, model changes and regression tests inside explicit repository contracts. Require observable evidence, not an agent's declaration that it works.

RUN AIVERIFY

Own the outcome

Review test results, deployment status, model invariants, rule freshness and recovery paths. A passing code change is not by itself proof of production behavior.

An important boundary

AI can help write and inspect the product. Retirement arithmetic belongs to deterministic, testable code — not an LLM inventing a forecast.

That separation makes the numbers reproducible. It also means an AI assistant can explain a result without being its financial source of truth.

One change, several places to verify

Use a fictional plan. Suppose a person records a shareholding that pays dividends, then chooses to reinvest those dividends instead of using them for monthly spending. A convincing interface might still show the income in two places. A dependable model must distinguish the economic event from spendable cash.

  1. Baseline: save the sample plan, including monthly spending, assets, assumed dividends and retirement age. Capture the same values on the overview and the asset page.
  2. Change one input: switch the fictional dividend from spendable to reinvested. Keep everything else fixed.
  3. Check the invariants: the dividend can influence invested capital, but it must not also appear as money available for that month's spending. Net worth and passive coverage should use their own declared definitions.
  4. Break the model: look for a duplicate transfer, a restricted account treated as immediately spendable, or an unchanged overview after an underlying account edit.
  5. Keep the evidence: record the initial assumptions, the exact change, calculated before/after values, any failure and the remaining limitation. Do not invent a success rate if no test was run.

This is a proposed check to reproduce, not a claim that these exact five steps were run as a benchmark for this note. The useful measure is whether the whole financial story remains consistent after a controlled change.

Privacy and portability are part of the engineering

At the time of this note, Bila Boleh offers browser-only planning by default, without connecting to a bank or brokerage. It also describes optional encrypted Sync for users who choose cross-device storage, alongside JSON export for portability. Browser-only copies are not encrypted at rest. The product's current privacy explanation describes the applicable choices and limits; this field note is not a security audit.

The trade-off matters. Saving in a browser is simple and private by default, but users can lose a device-only copy if browser storage is cleared. Sync adds recovery and security responsibilities. An export helps portability, but the file itself becomes sensitive. The interface needs to explain those differences rather than hiding them behind a Save button.

What still deserves scrutiny

  • Rules change. EPF, tax and other country-specific reference rules need effective dates and review, not an unlabelled permanent assumption.
  • Market numbers change. A user-entered portfolio value or forecast is not a real-time quote.
  • Financial projections are conditional. Inflation, dividends, spending and access rules can all differ from the scenario.
  • Model and product tests are different. A unit test cannot prove that saved inputs, import/export, browser privacy or a deployed page behave correctly.
What I am carrying forward

The real deliverable is not generated code. It is a question answered with traceable inputs, defensible rules, consistent state and checks another person can repeat.

This is a dated build reflection, not investment advice or an assertion that every possible financial edge case has been solved.

Read the 9order product case study →   Explore Bila Boleh →

Back to AI Field Guide →