FooFang Rhui

Context

Rakyat Fintech, from July 2026, live and in daily use. I am the only engineer on it: the server side, the search behind it, the incoming data, the cost tracking and the dashboard the team works from.

The problem

A small marketing team needs to know what competitors are doing, and needs a steady stream of adverts and posts. AI tools give you the second without the first: the writing comes out polished and convincing but connected to nothing that actually worked. Meanwhile the advertising budget is the one thing that drains quietly and cannot be got back. Two dangers had to be designed out before anything else was built. Spending money without a person choosing to, and the system learning from writing it made up itself.

Key decisions

1. Six separate steps, not one big instruction. Each step takes different information in and hands different work out. Asking one instruction to do all six blurs the joins where a person needs to step in, and those joins are exactly where approval belongs.
2. Nothing publishes or spends on its own. Every advert is created switched off, every image and video shows what it really cost in ringgit, and the whole process stops at a draft. The system is allowed to propose, and nothing more.
3. No practice data anywhere. Made-up test data would end up in the same store the system later reads from, quietly teaching it to trust its own inventions. Using only real data removes that route completely instead of trying to police it.
4. Two kinds of learning, kept apart. One looks at what was published and rewrites the plan: what to talk about. The other scores each finished video for how likely it is to spread, and takes one reusable filming technique from every example it studies: how to shoot it. Mixed together they produce advice too vague to use.

Where it stands

277 Python files, 112 screen components and 122 written design documents. Alongside the main process there is Studio, a chat where you make or fix one advert at a time. It follows the same brand rules, and when you ask for a change it works out whether you meant the words, the picture or both, rather than remaking everything.

The problem

After a break-in you can change a password. You cannot change a Malaysian IC number. It is issued once and identifies you for life. That changes who you are defending against. A stolen encrypted record is still worth having to someone who can only unlock it fifteen years from now. Normal encryption is a bet that the lock outlasts the value of what is inside, and here the value never runs out.

Key decisions

1. Two locks, not one. A new kind of encryption built to resist future quantum computers is combined with the well-proven kind used today, so the data stays private unless both are broken. The new method is young and not yet battle-tested; the old one is trusted but expected to fall to quantum computers eventually. Picking either alone bets everything on a single assumption.
2. The AI is never in the way of a live request. Four guards run constantly, blocking bulk data theft, spotting unusual behaviour, controlling what leaves the system, and overseeing the rest. All four are plain, predictable code. Anything that needs an AI to answer is something an attacker can slow down, confuse, or make expensive to run.
3. Personal details are removed before anything is stored, not after. A readable ID number is never written anywhere a later search could reach it.
4. It attacks itself, and the result is evidence rather than opinion. One AI tries to break into a sealed copy of the vault, but whether it succeeded is decided by a simple, unarguable check, not by the attacker’s own account. An AI marking its own work would fill the records with confident fiction.
5. A source that does not check out is treated as an error. In a security tool an invented reference is worse than no answer at all, so it fails loudly instead of quietly carrying on.

Outcome

No readable personal data is held anywhere in the system. The part that matters most runs with no AI involved at all: taking a file in, finding the sensitive details, scoring the risk, locking it and unlocking it. The critical path has nothing that can make things up.

The problem

Ask an AI why a share price moved and it will always give you an answer, usually a confident story about that one company, on a day the whole market happened to fall. The problem is not missing information. It is that polished writing looks exactly like real analysis by the time someone reads it, and where money is involved that gap is expensive. Anything supporting real decisions has to be able to say I cannot explain this one.

Key decisions

1. It refuses to pick shares, and says so plainly. No list of suggestions, no ranking. A system that both proposes ideas and argues for them will always find an argument – removing the proposing removes the temptation.
2. The numbers are separated before any story is told. Every price move is split into the part caused by the whole market, the part caused by that industry, and whatever is left over. It also reports how much it could not explain. That prevents the most common mistake in the field: blaming a company for a fall everybody had.
3. The limits sit outside the AI. How money is shared out, and five separate caps on how much of one thing you can hold, are ordinary code. Limits written as instructions to an AI are suggestions that clever wording wears down; limits written as arithmetic simply refuse.
4. Every prediction has a deadline and cannot be marked before it arrives. That keeps the scorecard honest, rather than a record of lucky guesses cashed in early.
5. The tests run with no internet and no accounts. A simple stand-in replaces the AI. Tests that call a real AI service are slow, unreliable and cost money, so people switch them off, and a test nobody runs protects nothing.

Outcome

At least 92% of the code must be covered by tests before any change is accepted. That runs automatically on both Windows and Linux and on two versions of Python, alongside style checks, type checks, deliberately hostile inputs and a start-up health check.