PodcastClipper AI

Inspiration

Short-form video is where podcasts get discovered — but inside a 60-minute episode, the 30 seconds worth clipping are buried. Finding them the old way means re-listening for hours. The clip tools that exist are built for English, cost money every month, and score your clips with a black box you can't question. As a creator working in both Bangla and English, I wanted a tool that treats non-English creators as first-class, costs nothing, and shows its work.

What it does

PodcastClipper AI turns a podcast transcript into a ready-to-post clip pack:

• Transcript input — Paste any Bangla or English transcript, or load the built-in 458-word Bangla sample in one click. (Minimum 30 words.)
• One-click analysis — Generates exactly 5 clip cards, each with an estimated episode timestamp, a hook title, a category badge (question, surprising statement, emotional peak) and a platform-fit badge for TikTok / Reels / Shorts.
• Transparent virality score — Every clip gets a 0-100 score ring with an expandable 'Why this score' breakdown. No black box, no fake 'AI magic' — honestly labeled as template-based in the UI.
• Ready-to-post assets — A caption with one-click copy and 8 hashtags per clip, plus a per-clip Markdown download.
• Sort & filter — Sort clips by score; filter by a minimum score threshold.
• Export pack — A 4-step launch checklist (with live progress) and one-click downloads of the full clip pack as TXT, Markdown, or CSV.
• History — Every analysis is saved in the browser: searchable, renameable, re-openable, deletable. No account needed.
• Bilingual UI — The entire interface, including generated captions and hashtags, switches instantly between Bangla and English.
• Mobile-ready — Fully responsive, verified down to 360px-wide phones.

How we built it

Built with Bolt.new (React 18 + TypeScript + Vite + Tailwind CSS). The app is 100% client-side — no backend, no accounts, no API keys — so it's fast and private by design. Clip detection uses a deterministic, rule-based scoring model: the transcript is segmented, scanned for hook patterns (questions, exclamations, contrast and emotion signals), and scored against published rules. Results persist in the browser's localStorage. The development process was spec-first: feature by feature, with a QA checklist run after each phase (final QA: 9/9 checks passed).

Challenges we ran into

• Explaining the score. A single unexplained number is useless. We rebuilt scoring as an explainable system where the breakdown must sum exactly to the displayed score — every point has a visible reason.
• Real bilingual support. Not just UI labels: generated captions and hashtags had to switch languages too, without breaking layout in either script.
• Zero backend, real persistence. Everything — including a full analysis history — had to work with no server at all, using only localStorage.

Accomplishments that we're proud of

• A genuinely finished product: every button works, exports download real files, nothing is a mockup or a placeholder.
• Scoring transparency that even paid tools don't offer — and honest 'template-based' labeling instead of pretending to be an AI black box.
• A Bangla-first tool that gives the world's 200M+ Bangla speakers the same creator tooling English speakers take for granted.

What we learned

Building with Bolt works best with a tight loop: write a precise spec, build one feature at a time, then test against a checklist before moving on. We also learned that honesty is a feature — users trust a score they can audit far more than one they're asked to believe.

What's next for PodcastClipper AI

• Real audio upload with automatic transcription
• One-click vertical clip rendering from timestamps
• Shareable public clip-pack pages for teams

Built With

• bolt.new
• localstorage
• react
• tailwind-css
• typescript
• vite