Bedir AI

Bedir AI is an educational answer product over Islamic primary sources. A question asked in natural language is answered from a curated collection of Qur'an verses and hadith, and the interface shows the exact passage each answer was drawn from, with a link to the source.

Company facts

Name	Bedir AI — operating and brand name
Identity	Bedir AI is operated by its founder, Mevlana Yalçın. Contractual and contact point: merhaba@bedirsavasi.com · Ankara, Türkiye.
Started	March 2026 · the domain `bedirsavasi.com` was registered on 8 March 2026, verifiable in the Verisign RDAP record.
Founder	Mevlana Yalçın — sole founder
Based in	Ankara, Türkiye
Team	1
Funding	Bootstrapped. No institutional investor, no accelerator, no outside round.
Contact	merhaba@bedirsavasi.com
Product	https://bedirsavasi.com/en/ (live, early access)
Infrastructure	Cloudflare Workers + Assets (site and answer endpoint) · Cloudflare KV (waitlist) · Cloudflare DNS and email routing

How the Claude API is used

The answer endpoint passes the corpus to the Claude API as text documents with server-side citations enabled, one document per verse or hadith, with a cache_control breakpoint on the last document so the collection is read from prompt cache on repeat questions. Claude is instructed to answer only from those documents, to reply in the language of the question, and to state when the current collection does not cover a topic rather than fall back on parametric knowledge. The server maps the returned document_index values back onto corpus entries and renders the quoted text plus a link to quran.com or sunnah.com.

All inference runs against the first-party Claude API through the Claude Console — not AWS Bedrock or Google Vertex AI. Anthropic is currently the only model provider in the product.

Backup engine, stated plainly

Answers are produced by claude-haiku-4-5-20251001 through the Claude Messages API: the collection is passed as document blocks with server-side citations and prompt caching enabled. Measured: a cold question writes a 64,610-token cache block and later questions read it back for roughly 130 fresh input tokens. The endpoint carries a daily token budget; when it is spent or Claude is unreachable, the answer comes from a labelled backup engine and both the provider field and the on-screen label say so - a backup answer is never presented as a Claude citation.

Policy posture

AI disclosure. The answer interface states that users are talking to an AI assistant and that answers are generated from cited sources and may contain errors.
No religious rulings. Bedir AI is an information tool. It does not issue fatwas and refers users to a qualified scholar when asked for one.
Inspectable sources. Every source text the product uses is published openly, with its own link to the publisher.

Verify this independently

Public repository, corpus and evaluation spec: github.com/mevlanayalcin/bedir-ai
Founder profile: linkedin.com/in/mevlanayalcin
Machine-readable source collection: /data/corpus.json
Domain registration record: bedirsavasi.com, created 2026-03-08 (Verisign RDAP)

Roadmap, with status

Shipped	Cited question answering over the Battle of Badr collection: 41 Qur'an entries, 70 hadith entries, in Turkish, English, German and Arabic
In progress	Expanding to the full Qur'an and the six canonical hadith collections; topic-level retrieval
Designed, not run	Published citation-accuracy evaluation: quotation fidelity and citation support scored separately (spec)
Next	A cited-answer API for Islamic schools and study platforms; mobile app

Figures on this page describe the source collection and the product, not audience size. We publish no user, traffic or revenue numbers because we do not have audited ones.