Dalilk Academy, 2026
AI Comment Replies
Every YouTube comment arrives sorted, with a fact-checked reply in the brand’s own voice.
At a glance
- comments collected
- 293K+
- replies drafted
- 6,905
- YouTube channels
- 6
- Role
- Product, design, full build
- Stack
- Python, Supabase + RLS, Next.js
Result
Since going live in late September 2026: 293,464 comments collected, 7,440 videos transcribed into a 35,000-item knowledge base, and 6,905 replies drafted for a 21-person team. 229 replies have been approved and sent since October 1. Nothing goes out without a person approving it.
The problem
Dalilk runs six YouTube channels. Comments pile up faster than the team can answer, and a plain AI reply can invent a price or a date. One wrong fact in public costs more than no reply at all.
What I built
- A worker collects new comments and transcribes each video with an in-house speech model.
- Comments are sorted into categories: question, complaint, praise, spam.
- A draft reply is written in the brand's voice, then a separate check verifies every fact against approved sources.
- The team reviews in an Arabic-first inbox and sends with one click through the YouTube API.
Stack
- Python
- Supabase + RLS
- Next.js
- TanStack Query
- YouTube Data API
Want something like this for your team?
Get in touch