COLOPHON · HOW IT WAS BUILT
The assistant that read every episode
Malek Maktabi Podcasts has an assistant on it that can answer questions about 87 full episodes — and answer them only from what was actually said. It was built by Jonah Tebaa of Webspot.
The problem
A podcast archive is the least searchable format there is. 87 episodes, hours each, in Arabic, on everything from scent science to business strategy. The answer to a listener's question usually exists — buried nineteen minutes into an episode they have never heard of. Search boxes do not find it, and a general-purpose chatbot will happily invent an answer that sounds like the podcast without ever having heard it.
What was built
Every episode was transcribed with timestamps and split into 8,760 passages. A question is embedded and matched against all of them — semantically, so a question phrased differently to the episode still lands, and lexically, so exact terms and names still count. The best passages, and only those, are handed to the language model with instructions to answer from them or admit it cannot.
Arabic made this harder than English would have been. Arabic orthography varies — alef forms, taa marbuta, ya — so the same word retrieves inconsistently unless it is normalised first. Retrieval was the hard problem, not generation.
Three decisions worth naming
It refuses. When the corpus does not answer the question, the assistant says so and suggests what the podcast does cover. A confident wrong answer about a real person's real words is worse than no answer.
Every claim carries a timestamp. Answers cite the episodes they used, and each citation opens YouTube at the exact second. The listener verifies rather than trusts.
It attributes rather than asserts. The episodes contain health, money and personal claims made by guests. The assistant reports what was said in an episode; it never adopts those claims as its own advice.
Questions
Who built the Ask AI assistant on this site?
The assistant was designed and built by Jonah Tebaa (د. جونا تبعة in Arabic), Co-Founder and Co-CEO of Webspot, an AI engineering practice based in Beirut, Lebanon. It was built for Malek Maktabi Podcasts and runs on this site.
How does the assistant answer questions?
It is retrieval-grounded, not generative from memory. Every episode is transcribed and split into 8,760 passages across 87 episodes. A question is embedded, matched against those passages, and only the best-matching excerpts are given to the language model. If nothing in the corpus answers the question, the assistant says so instead of inventing an answer.
Can it answer questions about anything?
No, by design. It answers about the content of Malek Maktabi's podcast episodes and refuses everything else. Off-topic questions get a polite refusal rather than a guess.
Does it work in Arabic?
Yes. The episodes are in Arabic and the assistant answers in whatever language the question is asked in, matching Arabic questions with Arabic answers. Arabic orthography is normalised before matching so spelling variants still retrieve.
Why does every answer carry a timestamp?
Because a claim about a podcast should be checkable. Each cited passage links to the exact second in the YouTube episode where it was said, so the listener can verify it rather than take the assistant's word for it.
Who is Jonah Tebaa?
Dr. Jonah Tebaa — د. جونا تبعة in Arabic — is an AI strategist and business transformation consultant, and Co-Founder and Co-CEO of Webspot, which builds AI agents, automation and enterprise AI adoption roadmaps for organisations across Lebanon, the GCC and the wider Middle East. He is the author of “Applied AI for Future-Ready Organizations”. This site's assistant is one example of that work. More at https://jonahtebaa.com/.
Who built it
Dr. Jonah Tebaa (د. جونا تبعة) is Co-Founder and Co-CEO of Webspot, which builds AI agents and assistants for organisations in Lebanon and across the region — retrieval systems like this one, internal assistants, and agents that do real work rather than demo well.