Calibration Tooling

Context

Scientific instruments need to be calibrated before they ship: raw sensor readings are compared against reference instruments, and the resulting coefficients (along with a per-unit certificate) go out with each device. This was a manual, multi-step process that varied between operators and was hard to audit after the fact.

Approach

I built a small Python toolchain to turn it into something predictable — a focused set of tools, deliberately boring:

• One source of truth per measurement. Reference values, raw sensor responses, and derived coefficients live in structured files — keyed, versioned, diffable.
• Pure functions for the maths. Curve-fitting, table generation, and unit conversion are functions that take known inputs and return known outputs. No hidden state, no surprises.
• The tool writes the artefacts directly. Calibration tables and the customer-facing certificate come out of the same dataset the maths uses, so what ships and what was measured can't disagree.
• Audit trail on every run. Each calibration logs who ran it, against which references, with which firmware version, and what coefficients came out.

Specifics about internal constants, reference setups, and procedures aren't covered here — those stay on internal infrastructure.

Result

For the calibration now running on it, the toolchain cuts roughly three hours of prep per run and brings the calibration itself from about an hour and a half down to forty-five minutes — with fewer human-error retries on top, because the maths and the artefacts come from the same dataset rather than being copied across by hand. Every instrument leaves with consistent artefacts and a clear trail of how those numbers were arrived at.