← All work

Film analysis · Traceable evidence

CineTrace

Connect film observations to traceable evidence.

Independent developer

FILM → TRACEABLE EVIDENCE

  1. 01

    Film source

    Source · keyframes

    Original media
  2. 02

    Shots & speech

    Intervals · transcripts

    Timestamp alignment
  3. 03

    Evidence report

    Analysis · references

    Trace observations

Keyframes · timestamps · evidence bundle

Analysis workflow illustration · not an application screenshot

01The project

A local-first film analysis tool that connects shots, speech and keyframes to traceable reports.

02The problem

Film analysis needs a way to trace observations back to the source material.

03Key decisions

Organize shot intervals, keyframes and speech transcripts into an evidence bundle; publish reports with source timestamps.

04Design in practice

Why traceability matters

A film analysis often moves between video playback, screenshots, transcripts and written notes. Once an observation enters a report, finding the precise moment behind it can become the slowest part of reviewing the argument.

CineTrace organizes that supporting material into a local evidence bundle. The goal is to let the reader return from an analytical claim to its timestamp and image, while the author remains responsible for interpretation.

A film becomes a navigable evidence bundle

Starting with a local video, the pipeline reads media metadata, detects visual intervals, extracts keyframes and calculates pacing statistics. Optional local speech recognition adds timestamped transcripts and confidence information.

Each run keeps its source fingerprint, tool and model versions, parameters, stage status and file checksums. Missing or failed stages remain visible, so a partial run cannot silently look like a complete analysis.

Publish a reading, with its evidence attached

The separate publishing module takes a Markdown analysis and resolves evidence anchors against a bundle. A timestamp can become an embedded frame; sequences can be shown as strips or contact sheets, alongside pacing and interval-duration charts.

The result is a self-contained HTML report that can be read outside the analysis environment. Growth charts require archived platform data; missing or invalid evidence stops publication instead of producing an unsupported illustration.

What I designed and built

I developed the Python pipeline, evidence organization and report publishing module, connecting FFmpeg, PySceneDetect and local speech recognition. The design keeps collection, verification and presentation as distinct steps.

A key decision was to preserve the difference between a detected interval, an edited shot and a narrative unit. Automatic boundaries help navigation, but they do not determine where a story beat begins or ends.

Verification and current boundaries

Bundle verification checks hashes, schemas, recalculated metrics and consistency between recorded status and produced files. It detects internal tampering and drift; proving correspondence with source media requires rerunning the analysis and comparing results.

This is currently a local command-line workflow with a separate report publisher. Optional model-generated visual observations remain review drafts, and absent platform data cannot support claims about reach or audience response.

05Outcome and current stage

A local analysis pipeline and an independent HTML report publishing module. Automated visual observations remain drafts for review.

Technical details & evidence

Role · Independent developer

  • Python
  • FFmpeg
  • PySceneDetect
  • faster-whisper