The problem
A curve is only useful if its result can be checked.
The measurement data contained noise and errors that affected the resulting curve. Reagent coefficients also needed to be accounted for. The numerical work therefore had two connected goals: improve accuracy and keep the calculation process practical.
The target was a sigmoidal representation of the curve. Work on smoothing and reconstruction had to be evaluated against calculated P50 and P90 values, rather than only by how convincing the curve looked on screen.
Numerical processing
Smoothing, reconstruction, and iteration tuning.
- Reduce measurement noise.Smooth the incoming measurements and correct errors affecting the resulting curve.
- Account for reagent coefficients.Incorporate these coefficients in the calculations.
- Investigate reconstruction methods.Explore Levenberg–Marquardt fitting as part of rebuilding a sigmoidal curve from smoothed measurements.
- Tune the calculation process.Adjust iteration counts to accelerate calculation while checking the resulting values.
Levenberg–Marquardt was an investigated fitting approach. The confirmed result is the numerical processing work and its validation; this case does not claim that the investigated optimizer became the final deployed method.
How we checked it
Several comparisons, one numerical result.
We compared the output with manual calculations, reference results, and other software. The numerical checks focused on deviations in the calculated P50 and P90 values. During research, visual overlays helped us examine how the curves differed.
P50 / P90
Compare deviations in calculated values.
Three sources
Manual calculations, reference results, and other software.
Curve overlays
Visually inspect the fit and investigate discrepancies.
Reduced errors, improved curve accuracy, and faster calculation. These are qualitative outcomes; no numerical accuracy percentage or runtime benchmark is asserted here.
Cloud platform & delivery
The numerical work was part of a larger research workflow.
I led a six-person engineering team and owned the cloud SaaS architecture and delivery. The platform supported automated device-data ingestion, cloud storage and backup, analytical dashboards, reports, CSV exports, audit trails, and remote research collaboration.
Device data
Automated ingestion, storage, and backup.
Measurements
Numerical processing and analytical dashboards.
Research results
Reports, CSV export, audit trails, and collaboration.
The web platform was deployed in the company’s research and validation environment. Its web and numerical-processing components contributed to a broader medical measurement process for which the company filed a patent application.
- Cloud SaaS
- Device-data ingestion
- Numerical processing
- Research collaboration
- Audit trails