How AI Works in Cuplikin: From Long Video to Ready-to-Post Clips

Many people imagine "AI in an app" as a single magic box: video goes in, viral clips come out. The reality is more down to earth, and precisely because it's down to earth, the pattern can be copied for other business needs. This article explains how AI works in Cuplikin step by step, in language that needs no technical background.

One note before we start: we're opening the concept, not the recipe. The prompts, how we judge moment quality, and the optimisation details are what we keep to ourselves. The concept itself isn't a secret, and understanding it will help you evaluate any AI project more clearly. For the story of how the product was built, read building Cuplikin from idea to release.

The big picture: five stages, one principle

Cuplikin's flow has five stages:

1. Video comes in, from a local file or a YouTube link.
2. Audio is transcribed into text with per-word timestamps.
3. AI finds and suggests moments worth clipping.
4. The user picks moments, tidies the transcript, sets captions and aspect ratio.
5. The clip is rendered with captions and saved to the device.

The principle behind it: AI suggests, humans decide. AI is used for the most tedious, time-consuming stage, while creative and final decisions stay with the user.

Stage 1: video comes in

The user picks a file from their computer or phone, or pastes a YouTube link. For links, App Store apps aren't allowed to run their own downloader tools, so the download is handled by a service on our server. The downloaded file is temporary: once it's copied to the user's device, the server copy is deleted automatically.

A small lesson here: platform constraints, like App Store sandbox rules, often shape architecture more strongly than technical preferences. We cover the impact in the case-study article.

Stage 2: on-device transcription

Audio is turned into text by Whisper, a speech-to-text model that runs directly on the user's device, not in the cloud. The output isn't just text, but every word with its start and end time.

Those per-word timestamps are the foundation of every later stage. With them, moments can be cut precisely, captions can appear word by word in sync with speech, and the user can click a word to jump to that part of the video.

Local transcription has two consequences we accepted knowingly: it needs a reasonably powerful device (the Mac version of Cuplikin supports Apple silicon only), and on phones the process is slower than on a server. The reward is privacy and predictable cost, which we discuss in the cost of running an AI product.

Stage 3: AI finds moments worth clipping

This is where the AI works. Its job is to find parts that can stand on their own as a clip: an engaging opening, complete content, and a natural ending. The result is a list of candidates, each with a time range, a suggested title, a score, and a short reason why that part was judged interesting.

Two things we consider important in designing this stage:

• The video is not uploaded for moment analysis. Heavy processing stays on the device, and only the minimum data needed is sent to the server. This protects privacy and keeps cost down at the same time.
• The AI must give reasons. A score without a reason is hard to trust and hard to correct. A short reason lets the user judge whether the AI "understood" the video or not.

Cuplikin offers two modes. Built-in AI uses our service and is counted in credits, which are only spent when the analysis succeeds. Custom mode lets users connect their own OpenAI-compatible provider, including local models through Ollama on Mac, so the entire flow can run without any data leaving the device at all.

Honest note: the AI can misjudge. Some moments the AI finds interesting fall flat when watched, and vice versa. We test quality by comparing the AI's suggestions against our own picks on a set of real videos, then refine the instructions repeatedly. No version is ever "finished". Anyone promising an AI that always picks the right moment is selling a dream. What's realistic is an AI that's right often enough to save time, with a human keeping the final say.

Stage 4: a human picks and tidies

The user sees the list of candidates, watches previews, and picks the ones to use. From there, the user can:

• Drag the start and end of the clip on the timeline.
• Correct the transcript text. Corrections update the captions immediately, because captions are built from the same transcript.
• Choose a caption template and video ratio (9:16 for Reels and Shorts, 1:1, 16:9, 4:3).

This stage is deliberately not automated. It's the part that makes the result feel like the user's own, and it's also the cheapest part computationally because it doesn't call the AI at all.

Stage 5: render and export on the device

The final clip is rendered with FFmpeg on the device: the video is cut, captions are "burned" into the picture, and the result is saved to a chosen folder on Mac or to the photo gallery on a phone, with no watermark. Projects and results never expire because they were never stored on our servers in the first place.

Patterns your business can reuse

If you strip away the "video" context, the flow above is a general pattern useful for many business needs:

Process first, then send to AI. Raw recordings are large and expensive to process in the cloud. Do the heavy work, like transcription, as close to the source as possible, then send the AI only the part it really needs for the task. For meeting recordings, customer-service calls, webinars, and interviews, a timestamped transcript is often enough: cheaper, faster, and easier to check.

AI suggests, humans decide. For almost every business case, this is the safest design. Meeting summaries approved by the minute-taker. Call quality scores confirmed by a supervisor. Content drafts edited by the marketing team. We discuss the reasoning further in the AI guide for small businesses.

Decide consciously what leaves the device or the company. In Cuplikin, the video isn't uploaded for moment analysis; only the minimum data needed leaves the device. Businesses with sensitive data can use the same pattern: process as much as possible inside, send out as little as possible, or use a self-hosted model. We cover this choice in RAG for business documents.

Count usage per success. Tying cost to work that succeeded, not to attempts, builds user trust and makes cost predictable.

Some applications of this pattern we've discussed with prospective clients:

Next steps

If your business has recordings, documents, or conversations piling up and you want them processed with a pattern like this, we can help design the flow, including deciding what's processed locally and what goes to the cloud. Start with a free consultation or chat on WhatsApp. Our AI engineering services are on the services page.