Assalamu alaikum wa rahmatullahi wa barakatuh.
I am writing to you from a position I did not expect to be in: building infrastructure for a problem that has no good solution yet.
The problem is this. Islamic AI is here. Muslims are using AI assistants to learn about their deen, to ask about halal permissibility, to look up hadith, to understand fiqh rulings. Islamic edtech companies are shipping products with AI features. Quran apps are adding generative explanations. Masjid software vendors are adding Islamic Q&A. The builders doing this work are, in many cases, Muslim builders who care deeply about what they are making.
And the AI systems they are building on top of were not trained to be right about Islam. They were trained to sound right. Those are not the same thing.
What I kept seeing
I kept seeing the same pattern. A well-meaning Islamic AI product ships. A Muslim user asks about a hadith. The product returns a confident answer with a source citation. The citation is wrong. Not obviously wrong. Subtly wrong in a way that requires checking the Bukhari collection yourself to catch. A young Muslim who trusts the product never checks.
I saw a Quran app add AI-generated tafsir. The tafsir cited a verse with the correct surah name but the wrong ayah number. The meaning was close enough that most users would not notice. Some would notice and lose trust in the app. Most would not notice and build their understanding on a wrong foundation.
I saw an Islamic finance tool provide a fiqh ruling on a specific transaction structure. The ruling was accurate for the Hanafi position. It was not accurate for the Shafii position, which is what the user in the conversation had indicated they follow. The tool had no way to know this, and no mechanism to say so.
These are not edge cases. They are the default behavior of AI systems applied to Islamic content without a verification layer. The default behavior is confident and wrong at a rate that is unacceptable for content that shapes how Muslims live their deen.
Why I built Tasfi
I built Tasfi because I could not find the thing I needed. I looked for an Islamic AI verification API. There was not one. I looked for a public benchmark of Islamic AI accuracy. There was not one. I looked for a checksummed, licensed source bundle that I could point a verifier at and say "check against this." There was not one.
So I built the source bundle. I built the verifier. I built the benchmark. I built the Trust Receipt format that gives Muslim builders a paper trail they can show their users, their scholars, and their own nafs.
I am one person. I am not a scholar. The boundary is explicit on every Tasfi Trust Receipt: not a fatwa service, not a scholar replacement. I cannot be those things and I am not trying to be. What I can build is the engineering layer that clears the path of mechanical errors so that scholars can focus on the questions that require their knowledge.
What Tasfi asks of you
If you are building an Islamic AI product, Tasfi asks one thing: put verification between your model and your Muslim users.
Not because the verification is perfect. It is not. Tasfi Guard catches what the approved source bundle can verify and stops there. Novel questions, contested scholarly matters, madhab rulings that are not addressed in the methodology documents: Guard does not adjudicate those. The receipt says so.
But the errors that Guard does catch are real. Fabricated Quran citations. Mis-attributed hadith. Ungraded narrations presented as sahih. Rulings attributed to all four madhabs that are only the position of one. These are the errors that the open web is full of, that AI systems reproduce confidently, and that a verification layer can catch before a Muslim user sees them.
If you put Guard in your stack, you are not claiming perfection. You are claiming due diligence. You are saying: we checked what could be checked. We have a receipt. Here is the paper trail. Our scholars can review the flags. Our users can verify the receipts. We did not ship without looking.
That is the ask. Not perfection. Due diligence, with a receipt.
What Tasfi promises in return
Tasfi promises to stay inside its boundary. It will not issue fatawa. It will not replace scholars. It will not claim to be a religious authority. Every Trust Receipt will say this explicitly, for as long as Tasfi exists.
Tasfi promises to keep the source bundle honest. Every source in the approved registry has a license, a provenance, a checksum, and a lifecycle state. When a source changes, the bundle version changes. When a source is removed, the receipts from before the removal remain valid for their source bundle version. The source changelog is public.
Tasfi promises to keep the benchmark open. The 420-fixture public suite is CC-BY 4.0. Any Islamic AI verifier team can run their system against it and submit results. The methodology is documented. The leaderboard is public. If Tasfi Guard is outperformed on the benchmark by another verifier, that information will be on the leaderboard where everyone can see it.
Tasfi promises to build toward the scholar layer. The engineering is ready to support a signed scholar review workflow: a named advisor who reviews the methodology, approves source additions, and signs off on scope expansions. Building that relationship is the next organizational step. When it exists, you will see a named scholar on the methodology page.
The amanah dimension
I want to say one more thing about why this matters beyond the product.
The infrastructure that decides which Islamic answers reach Muslims is amanah. It is not neutral technology. It is a trust layer with a religious dimension. The Quran, the sahih hadith corpus, and the four Sunni madhabs are amanah. The system that gates AI access to them is also amanah.
I did not build Tasfi to exclude anyone. Non-Muslim builders are welcome to integrate Guard as technical infrastructure. The API is open. The benchmark is open. The code is Apache-licensed. But the standards layer, the source registry, the methodology, the boundary enforcement: those belong with the Ummah, built by Muslims, accountable to the community whose deen is at stake.
That is not a political claim. It is a stewardship claim. The Quran was preserved by the Ummah. The hadith were authenticated by the Ummah's scholars. The trust layer for AI access to those sources should be owned, built, and governed by people who understand that this is not merely a technical problem. It is an obligation.
Wa billahi tawfiq.
Kevin Richards
Founder, Tasfi Enterprises
founder@tasfi.app