KITQU

The product position was fixed early: a workflow layer, not a text generator. That has a real cost. Every feature whose only effect is producing paragraphs faster gets rejected, even when it is the easiest thing to sell.

Cost is the second limit. Pricing is token-based, so the cost of a request has to be known before the bill arrives, not after. Without that, pricing is a guess.

The third limit is dependence on a single model provider. Prices and availability change outside our control, and locking in means handing the product's cost structure to someone else.

Finally, the infrastructure runs on a self-hosted VPS on a small team's budget — there is no room to train or fine-tune a model of our own.

I built a multi-provider AI router behind one internal interface. Feature code calls the router, never a provider directly, so switching providers never touches feature code. The router is also where token telemetry lives, so the cost of every request is recorded in one place.

Revision memory is stored as persistent state — we call it Catatan Bimbingan — rather than stuffed into the context window. Context windows run out and cost money; stored state can be re-read, edited by the student, and survives across sessions.

Supervisor style calibration sits in a layer separate from the writing itself, so one draft can be adapted to a different supervisor without being rewritten.

The AI defence simulation has three modes — Santai, Realistis, Killer — because a single difficulty level is useless: too soft trains nothing, too harsh makes people stop using it.

The trade-off I rejected: training or fine-tuning a model of our own. The cost does not justify it, and the actual problem lives in the workflow, not in model quality.

Distribution is automated through a Telegram bot that posts to Instagram via the Graph API, so marketing does not require another person.