Case Study · AI Automation · Aviation
AI-Powered Aircraft Performance Chart Digitization
An AI Chart Simulator that converts performance chart PDFs into the finished Excel workbook, complete with the verification overlay
The result: a graph that currently takes one to two hours can be completed and reviewed in a few minutes, with about 95% of charts expected to complete cleanly.
The Customer
A flight-planning application built on aircraft performance data
Client X requires aircraft performance charts, such as takeoff and landing distance, rate of climb and cruise performance, to be converted from manufacturer PDFs into structured numeric tables that feed its flight-planning application.
At a glance
- Customer
- Provider of a flight-planning application
- Charts
- Takeoff and landing distance, rate of climb, cruise performance
- Volume
- Around 1,000 graphs a month
- Manual effort
- About 15 analysts, one to two hours per graph
- Solution
- AI Chart Simulator web application
- Built on
- Python, Claude API (vision), Microsoft Graph API (OneDrive)
The Problem
About 1,000 graphs a month, digitized by hand, point by point
This conversion is currently performed manually at an outsourced vendor, where a team of approximately 15 analysts digitizes around 1,000 graphs a month, point by point, in WebPlotDigitizer. Each graph takes between one and two hours depending on its complexity. The process is slow, its accuracy depends on the individual analyst, and capacity can only be increased by adding headcount.
The current process
- 1
Intake. Chart PDFs arrive from Client X by email and are assigned to analysts manually.
- 2
Calibrate. For each graph, the analyst loads the chart into WebPlotDigitizer and calibrates the axes by selecting four reference points and entering their printed values.
- 3
Plot. Each curve is then plotted point by point, beginning with the lowest line, with a point at every major gridline crossing, additional points along each bend, and a final point where the line ends.
- 4
Export. The coordinates are exported to the Excel template and rounded to the client's standard using MROUND.
- 5
Verify. To verify the result, the analyst plots the data in Excel, overlays it on the original chart image at up to 400% zoom, aligns the gridlines, and confirms that every plotted line sits on its printed counterpart.
- 6
QC. A senior analyst then repeats this overlay check as QC before the workbook is delivered to the client.
Pain points
- Each graph consumes one to two hours of skilled effort, and output quality depends on the individual analyst rather than a repeatable process.
- Curves that cross or fan out closely are the most common source of errors and rework, and many source charts are low-resolution scans in which gridlines and curves are difficult to distinguish.
- Because QC repeats the full overlay check, the verification effort is incurred twice.
- No work is reusable, so every chart is digitized from scratch.
The Results
What AI chart digitization changes
A few minutes
to complete and review a graph that takes one to two hours by hand
~95%
of charts expected to complete cleanly, with the remainder flagged for an analyst to correct
Built in
verification overlay embedded in every workbook, in the existing analyst format
The AI Automation Solution
An AI Chart Simulator that works the way the most experienced analysts do
The AI Chart Simulator is a web application that converts a performance chart PDF into the finished Excel workbook, complete with the verification overlay analysts use to check their work. The output follows the existing analyst format exactly, so no downstream process needs to change.
The application follows the same method as the most experienced analysts: it calibrates the axes against the printed labels, counts the lines on the chart before tracing them, and places points at gridline crossings, bends and endpoints. The extracted values are then plotted back onto the original image for visual verification, and the application never reports a line that is not printed on the chart.
It is designed as a first-pass automation with human review rather than a fully unattended system. Approximately 95% of charts are expected to complete cleanly, with the remainder flagged for an analyst to correct. A graph that currently takes one to two hours can be completed and reviewed in a few minutes.
Every AI run, in four steps
- STEP 1Axes and grid calibration
- STEP 2Curve tracing
- STEP 3Dot plotting
- STEP 4Line grouping
Each step can be inspected and corrected, alongside a results table in which every value maps back to a point on the chart.
See the AI Chart Simulator in Action
1. Intake.
On the Intake screen, users run chart PDFs from a shared OneDrive folder or upload them directly.
2. Runs.
Each run on the Runs screen is broken into four steps (axes and grid calibration, curve tracing, dot plotting and line grouping), each of which can be inspected and corrected, alongside a results table in which every value maps back to a point on the chart.
3. Output.
The Output screen holds the finished Excel workbooks, which can also be exported as CSV or as an overlay image.
4. Cost Consumption.
The Cost Consumption screen shows the Claude API spend for each run.
App Description
AI that stops when it is not confident
The Chart Simulator is a web application organized around four screens: Intake, Runs, Output and Cost Consumption.
The application stops when it is not confident. If axis calibration deviates by more than two pixels, the run halts rather than producing inaccurate values, and any traced line that is not supported by printed ink is discarded. Charts that cannot be completed cleanly are flagged for analyst review.
The application is built on a Python processing pipeline with the Claude API as its vision engine, and connects to OneDrive through the Microsoft Graph API. Rendering, axis calibration, re-plotting and Excel generation are handled in code, while Claude reads the chart structure and curve paths, with calibration checks, ink verification and human review around it.
Key features
- Imports chart PDFs from a shared OneDrive folder or by direct upload
- Produces the finished Excel workbook in the existing analyst format
- Embeds the verification overlay in each workbook automatically
- Breaks each run into four steps that can be inspected and corrected
- Allows values to be edited in the results table, keeping the overlay and workbook in sync
- Halts when axis calibration is inaccurate and never reports a line that is not printed on the chart
- Flags charts that cannot be completed cleanly for analyst review
- Handles both digital PDFs and low-quality scans
- Exports results as Excel, CSV or an overlay image
- Shows the Claude API cost of each run
- Reduces a one-to-two-hour manual task to a few minutes
Get In Touch