It will tell you when it doesn’t know
Every answer carries a confidence score. Below the threshold it abstains rather than guesses — because on matters of Deen, “confidently wrong” is the worst possible outcome.
Built to be trusted with Deen
Most assistants optimise for always having an answer. This one optimises for never giving you a wrong one.
Doubtful rulings go to a scholar
Religious answers that score low are queued for human review. A scholar approves, corrects, or rejects them — the verdict is recorded, not discarded.
Citations are verified, not decorative
Qur’an and hadith references come back as typed, bounds-checked objects — validated against the 114-surah index and matched to corpus text, not trusted because the model wrote them.
How an answer is made
Four gates between your question and the reply
1
Your question is gated
The input passes a policy check before a single token is generated.
2
Sources are retrieved
Relevant tafsir and references are pulled from named works and handed to the model as grounding.
3
The answer is scored
Citations are verified, hadith are graded, and a confidence score decides: answer, hedge, or abstain.
4
You get it, or a scholar does
High confidence reaches you directly. Low-confidence religious answers go to the review queue first.
What it can do
Not a general chatbot with a prompt taped on — purpose-built pieces, each solving a way an Islamic assistant can fail
Tafsir from named works
Ayah explanations retrieved from classical tafsir — al-Muyassar, Fi Zilal al-Qur’an, Ma‘arif al-Qur’an, Bayan ul Quran — with attribution. Never paraphrased from model memory.
Grading-aware hadith
References across Bukhari, Muslim, Abu Dawud, at-Tirmidhi, an-Nasa’i, Ibn Majah, and the Muwatta are graded before they reach you — a da‘if narration is never dressed up as sahih.
Madhhab-aware fiqh
Fiqh questions are classified and answered with the position of the Hanafi, Maliki, Shafi‘i, and Hanbali schools rather than flattened into one opinion.
Answers in your language
Arabic, English, Urdu, Malay, French and more — auto-detected or requested. Qur’an is always quoted in Arabic script with a translation and a surah:ayah reference.
Remembers, on your terms
Builds a profile of your knowledge level, madhhab, and what you’ve studied, carried across sessions. You can read it, delete it outright, or opt out per message.
Live zakat and nisab
Nisab is derived from the current gold price rather than a stale constant, and zakat can be calculated against the USDC balance in your own Stellar wallet.
Quizzes and flashcards
Turn a topic or a lesson into schema-validated study material at three difficulty levels, so revision isn’t another thing you have to build yourself.
Layered safety
A policy pipeline gates both what goes in and what comes out, so the assistant stays inside the boundaries it was given.
Your memory, your call
What the assistant remembers about you is readable on request and erasable in full. You can also switch remembering off for any single message. Stored profiles expire on their own — they are not kept indefinitely.
Open source, end to end
The assistant is a FastAPI service under an MIT licence, built in public alongside the web app and the API. If you want to know exactly how a ruling is scored or a citation is checked, you can go and read it.
github.com/Deen-Bridge/dnb-ai