All work
AI & automationOpen source

Video Verification Pipeline

A Python pipeline that turns a screen recording of a verification page into structured, auditable JSON: OpenCV frames, parallel OCR and an optional LLM structuring step.

Product
AI document pipeline
Built with
Python, OpenCV, Nanonets
The five pipeline stages, from input video to structured JSON
5

inspectable stages, each saved to disk

Parallel

OCR, one worker per frame

100%

of fields traceable to their source frame

Overview

The problem we set out to solve.

It takes one screen recording of someone opening a licence lookup or certificate page and turns it into structured JSON that can be audited. Browser text is tiny in video, so the work is split into small steps instead of asking one model to do everything.

  • Browser text in a video is small, compressed and hard to read.
  • Verification results must be auditable, not a black-box guess.
  • The LLM must never become the source of truth.
Every frame read at the same time by a thread pool
Raw OCR in, one clean structured record out
What we built

Every part, built to work in the real world.

Frame extraction

OpenCV samples the video, by default one frame every three seconds.

Preprocessing

Grayscale, 1.5× upscale, adaptive threshold and sharpening make small text readable.

Parallel OCR

Every frame goes to Nanonets OCR at once, then results are sorted back into order.

Optional structuring

Groq formats the OCR into a strict schema; anything not found stays null.

CLI and API

Runs from the command line or through a FastAPI upload endpoint.

Full audit trail

Frames, raw OCR and structured records are all written to disk.

How it's built

Every stage writes its result to disk, so a reviewer can trace any field back to the frame it came from. OCR is the source of truth; the model only formats, and it can be switched off in one setting.

  • PythonPipeline and CLI
  • OpenCVFrames and preprocessing
  • Nanonets OCRParallel text extraction
  • GroqOptional structuring step
  • FastAPIUpload-and-verify API
The outcome

A reviewable verification workflow for video: every intermediate result is on disk, OCR runs in parallel, and the LLM step is strictly optional.

Prefolk landing page: “Your first users. Your blind spots. Before launch.”
Next project

Prefolk

AI user-testing SaaS

Have an idea? Let's build it.

Tell us what you want to build. You'll get a clear plan, a timeline and a fixed price within a week.