Surauanswers you can check

the brief

Surau didn't begin with a technology. It began with a discomfort. Ask an AI a question about Islam and it answers instantly, fluently, with total confidence. Ask where the answer came from, and the confidence falls apart. For a recipe that's fine. For someone's faith it's quietly dangerous: a person can be confidently misled about something they hold sacred, and never know it.

So the product got one job, and it became my entire brief: never let someone believe an answer they can't verify.

designing for verification

Most products in this space optimise for more answers, faster. I chose the opposite. A single answer you can trace back to a real page in a real book is worth more than a thousand fluent answers you have to take on faith.

I designed the conversation like an interface. Where does a reader get confused? Where do they need proof instead of prose? Every answer hands the reader the page: the heading, the exact page number, a verbatim quote and a link straight into the source.

designing for refusal

Everyone's instinct is to make the product answer better. I spent my energy on the opposite: making it trustworthy enough to say "this isn't in the source." A product about faith that confidently invents an answer is worse than one that admits a limit, so a graceful, honest "I don't know" was a deliberate design decision.

ai in the loop

Technology can organise knowledge and enforce a process. It can't judge whether a translation is faithful to a sacred source. That judgment belongs to people with real training.

1. classical texts
2. retrieval
3. draft → review → publish

step	the agent	me
answers	finds the passage and drafts the answer	designed every answer to carry the heading, the page, a verbatim quote and a link
limits	says when a question isn't in the source	made that refusal feel like integrity, not failure
review	enforces draft, review, publish, so nothing reaches a reader unreviewed	brought in a scholar from Al-Azhar University to judge whether what's reviewed is right
scale	ran retrieval that genuinely worked	did the math at scale, and paused the costly part

the hard calls

Partway through, I watched a process quietly burn far more resources than it should, saw the warning, and let it run anyway. I paid roughly twelve times over for that lesson: a negative signal is a reason to stop.

Then the hardest call. Retrieval worked, but as a solo builder the economics at scale didn't close, so I paused the costly part and changed the order: build the clean, verified library first, starting with the Qur'an on an official, government-published foundation, and earn the right to retrieve over it later.

where it's headed

A clean, canonical, human-verified library of classical Islamic knowledge, and only then answers built on data that has earned them. Shipping the wrong thing fast is still shipping the wrong thing.