A small web app that returns verified verses of Rumi instead of AI-generated ones, designed to protect a literary heritage from misattribution and hallucination.
The Challenge
Before any code, I tested an early concept with two colleagues unfamiliar with Rumi — one in the Netherlands, one with a Japanese cultural background in Australia. Their first reaction: “I'm not clear what the solution will be… why do you care?”
Newcomers needed the purpose in seconds. Readers who know the work needed accurate sourcing before they would trust it at all.
The obvious version lets a model write a “Rumi-style” reflection on demand. That is also the fastest way to fabricate the heritage the product exists to protect.
Japanese, unfamiliar with Rumi. Needs to understand what the product is and why it matters within seconds.
International, deeply familiar with Rumi. Needs accurate sourcing and interpretation before trusting the experience.
Shared need: Credibility must be visible, not merely claimed.
The real challenge wasn't building an AI feature. It was designing constraints that kept the AI from ever inventing text on a literary heritage it wasn't authorized to speak for.
My Approach
Reframing the idea around ethical AI turned confusion into clarity. The same reviewer went from “why do you care?” to “This I get. This is crystal clear.” — Vlad, concept reviewer
A Rumi study group (n=4) was asked how they find verses and whether they had doubted a quote online. 3 of 4 had. The trust problem was not hypothetical.
Four rules the product would not break: retrieve, never generate; public-domain sourcing only; human-authored reflection; a structured emotional taxonomy instead of an open prompt.
Results from the Rumi study group survey (n=4) referenced in Step 02.
Key Decisions
A model that composes “Rumi-style” text will eventually present something Rumi never wrote, with full confidence. Restricting the system to a fixed, curated set removes that failure mode.
Free-text emotional input felt arbitrary and read like a generic chatbot. Ten emotional categories give a clear starting point and keep matching constrained.
Popular English translations of Rumi carry copyright restrictions. Reynold A. Nicholson's Mathnavi is public domain, so the product can be shared without adding another unclear source.
People also want to check quotes they have already seen. Verify returns one of five outcomes: Confirmed, Confirmed partial, Confirmed multiple, Close but not exact, or Not found.
The biggest challenge wasn't making the product feel intelligent. It was making sure it never pretended to know something it didn't.
What Broke
A “jealousy” selection once returned an unrelated verse. Left unfixed, that quietly breaks the promise that what you see is the verse you asked for.
Testers treated Verify as a search bar and typed single words such as “water.” That is a framing problem, not a logic problem.
A professional designer reviewing the live product flagged the logo as too small. It was fixed the same day.
Self-testing caught an altered quote still marked Confirmed. It was fixed before any real user saw it.
“This is awesome Delara! …this can be a great and more responsible way to use AI.”
“The value: ensuring AI is not hallucinating… Then I can see that this is highly accurate translations.”
— Odette
“I found the translation comprehensive… it captures the meanings very well.”
— reader from a Rumi-interest group
Evaluation
Before anyone else touched the product, adversarial self-testing caught three issues a casual pass would miss: an altered quote, one word changed so the meaning reversed, still marked Confirmed; a valid partial line returned as not found; and a phrase that appears in several verses returned only the first match. All three were fixed and re-checked.
Real use then surfaced what self-testing could not: the category-verse mismatch, Verify used as search, and the logo note from a designer. Early signals are intentionally modest: 92 category views, 17 verse ratings, 100% helpful ratings, and 5 verification attempts.
Final Design
Ten emotional categories replace free-text input, so matching starts from a clear choice rather than an open prompt.
Every verse is shown with its exact book and line. The reflection is human-authored, not generated by the model.
Paste a quote attributed to Rumi and check it against Books 1–4 of the Mathnavi. The system states what it found, including when it found nothing.
Reflection
This round relied on 4 survey respondents and a handful of testers, evaluated by me alone. An independent reviewer would reduce the bias of my own closeness to the material.
The set covers 30 verses across Books 1 to 4 of the Mathnavi only. Wider coverage would make it useful beyond a focused demo.
More than one tester treated Verify like a search bar. That is the clearest next fix, and it is a copy problem, not a matching problem.
Verse translations are Reynold A. Nicholson's public-domain English translation of the Mathnavi, Books 1–4. Interpretations are Persian-language summaries of Parviz Shahbazi's teachings on the Ganj-e Hozour program, translated into English by the author. Trust is designed into the content model, not added as a disclaimer.
Generative AI (Claude, Anthropic) supported technical architecture, application code, wording, and the structure of the research log. Verse selection, citations, interpretations, and authenticity verdicts were reviewed by me for accuracy.