TraceAI

Inspiration## Inspiration

Public visual search often returns fragmented links, direct image files, duplicated results, weak context, and uncertain matches. A reviewer may find a potentially relevant image but still lack a clear source page, evidence trail, or reliable way to preserve the review.

TraceAI was created to turn those scattered results into structured, source-linked visual evidence leads that a person can review safely.

What it does

TraceAI allows a reviewer to upload one or more reference photographs and run a bounded search across public visual sources.

The platform:

• collects possible visual candidates from public sources;
• groups results into structured evidence cards;
• preserves source URLs and available context;
• shows which search routes were checked;
• marks candidates that are not independently confirmed;
• separates user-facing evidence from technical diagnostics;
• preserves recoverable search state after refresh or connection loss;
• requires human review before any conclusion is made.

TraceAI does not automatically determine identity, ownership, infringement, unlawful use, or takedown eligibility.

How we built it

TraceAI is a Python and Flask web application with a JavaScript interface, a SQLite-backed review layer, background public-discovery processing, and a production deployment using Gunicorn, Nginx, and Oracle Cloud Infrastructure.

During OpenAI Build Week, I used Codex with GPT-5.6 Sol to repair and extend the end-to-end workflow.

The Build Week work included:

• recoverable server-side search state;
• restoration of results after browser refresh or connection loss;
• privacy-safe persistence of review results;
• regression coverage for the user-facing report;
• validation of result separation and recovery behaviour;
• production deployment of the tested changes.

The focused recovery and report test suite passed 105 tests. The changes were committed and pushed to the project repository.

Codex feedback session:

019f8419-950a-7890-93ab-27bd9d14e76b

Challenges

The main challenge was not simply finding images. It was preserving enough source context to make each result useful for human review while avoiding unsupported conclusions.

Other challenges included:

• handling slow or incomplete public-source searches;
• recovering state after network interruption;
• preventing raw diagnostics from appearing in user results;
• deduplicating candidates from multiple routes;
• clearly distinguishing possible leads from confirmed evidence;
• keeping the workflow usable within a practical latency budget.

What I learned

I learned that a trustworthy visual-search workflow needs more than similarity scoring. Reviewers need source context, uncertainty labels, stable records, recovery after interruption, and a clear separation between machine-generated signals and human conclusions.

Codex with GPT-5.6 Sol was especially useful for tracing the existing workflow across backend, frontend, tests, and deployment code, then implementing a focused fix without replacing the underlying product architecture.

What's next

The next steps are to expand public-source acquisition, improve source-page recovery, strengthen evidence packaging, and continue building a larger privacy-conscious visual index.

TraceAI is developed by Facewipe LTD.

UK patent pending — application GB2612078.2.

Built With

• apis
• cloud
• codex
• computer
• css
• flask
• git
• github
• gpt-5.6
• gunicorn
• html
• javascript
• nginx
• openai
• oracle
• python
• rest
• sqlite
• vision