TraceParse Verify

Inspiration

Engineers and researchers often need numerical data from graphs in papers, reports, and legacy documents. Manual digitization can be slow, while automatic tools may return data without making it clear whether the extracted curve actually matches the source.

TraceParse Verify was created to make graph-to-CSV extraction easier to review and trust.

What it does

Users upload a PNG or JPG graph, mark the plot boundaries, enter the linear-axis ranges, and select one representative pixel from the coloured curve.

TraceParse Verify then:

extracts the selected curve;
converts image coordinates into numerical x-y values;
draws the extracted trace directly over the original graph;
reports coverage, longest gap, point count, fragmentation, and ambiguity warnings;
exports the calibrated values as CSV.

The workflow keeps the important calibration and verification decisions visible to the user.

How I built it

Before Build Week, I used Claude to explore an early prototype for PDF rendering and colour-based graph tracing.

During Build Week, I used Codex with GPT-5.6 Sol to generalize the extraction beyond hardcoded image dimensions and build the TraceParse Verify Streamlit application. Codex also helped add interactive calibration, visual overlay verification, deterministic quality metrics, automated tests, documentation, and the GitHub submission workflow.

The application is built with Python, Streamlit, OpenCV, NumPy, and Pandas.

Challenges

The main challenge was that scientific graphs often contain multiple curves, similar colours, anti-aliasing, labels, and incomplete visible traces.

The current MVP therefore focuses on one non-black coloured curve with linear axes. It uses adjustable colour tolerance and quality warnings so uncertain results are not silently presented as correct.

Another challenge was accurate manual calibration of the plot boundaries. A zoom-assisted calibration tool is planned as a future improvement.

Accomplishments

Built a working Streamlit interface.
Generalized the tracer to support different image dimensions.
Added visual overlay verification.
Added CSV export and deterministic quality metrics.
Added 16 automated tests.
Preserved a clear separation between the earlier prototype and the new Build Week implementation.
Created a reproducible sample graph with known reference data.

What I learned

I learned how important human verification is when converting visual engineering evidence into numerical data. A useful extraction tool should not only generate a CSV, but also show the user why the result may or may not be trustworthy.