%

Media / AI

Case study

Impact*

An AI-powered video processing API that detects high-impact moments in fight videos and applies dramatic visual effects automatically.

Project

Impact

Industry

Media / AI

Year

2025 - 2026

Role

Backend Engineer

The problem

Impact frames come from anime. When a big hit lands, the picture snaps to stark, high-contrast black and white for a split second, and the hit lands twice as hard. Fight edits borrow the trick all the time, but adding them by hand means hunting down the exact frame of every hit and cutting the effect in yourself.

How it works

Impact is a Go API that takes a video and returns it with the effect already applied. One request does the whole thing.

  • FFmpeg pulls frames at 10 per second, scales them down and burns a frame number into each one.
  • The frames are tiled into 5 by 4 grids, so a 30 second clip goes to Gemini as about 15 images instead of 300. The numbers on each frame let the model say exactly which frames it means.
  • The grids are uploaded five at a time, with retries and backoff.
  • The prompt spells out what counts as real contact and what doesn't, such as wind-ups, aftermath and transitions, and asks for the best two to five moments as frame numbers.
  • Each frame number becomes a short time window. Overlapping windows are merged, and FFmpeg applies a high-contrast black and white flash during those windows.

The result

Upload a fight clip and get it back with the flashes on the hits. The response also lists the detected frames and what the Gemini calls cost, worked out from a per-model price table.

KAKCIL

Democratic AI platform that routes queries across multiple LLMs, enabling real-time inter-model debate and consensus voting for high-confidence responses.

Next project

AI