Engineering platform
Artyc
A staged push any engineer can run. Pick the build, pick the stage, put a name and a reason on it. The next stage opens when data arrives from the new image and not when a shipper says it updated.
A staged push any engineer can run. Pick the build, pick the stage, put a name and a reason on it. The next stage opens when data arrives from the new image and not when a shipper says it updated.
The short version
Artyc makes battery-powered coolers for medical transport that hold temperature for days without plugging into anything. Proving that means chamber runs that bake and freeze a unit for hours, calibration against a reference probe, board tests on the manufacturing line, and firmware going out to units already in the field.
The mechanical, firmware, and software engineers performing this work had workflows running in disparate places, none of which produced anything the rest of the company could easily read.
A platform underneath all of it was missing. Run the thing, record what happened, hand it to somebody else, and still be able to find it a year later. We needed consolidation and tidiness because not having results readily available is as good as not having results at all.
My role was to work across the engineering disciplines that don't always share a vocabulary. A mechanical engineer describes a run by what the chamber did. A firmware engineer describes the same run by what the unit reported. Getting those to be the same record was most of the job.
Firmware became a staged, gated rollout that any engineer can run, not just the one who wrote it. Thermal verdicts are computed from the statistics on the page. Calibration accepts outside lab evidence against the unit itself. Board tests no longer have a two-state verdict, so a skipped test no longer counts as a pass.
Eleven sensors, eight hours. Verdicts compute from per-sensor statistics. Evidence stays visible so a later reviewer does not have to retest to disagree.
The problem
Each team had its own version of the same failure.
Thermal verification generates good data and no way to pass it on. Whoever ran the chamber had the answer and everybody else had to ask them for it.
Calibrations are performed by a third party, so the evidence that a shipper reads true arrives from outside the company. Certificates landed in email with no object to attach them to.
Board tests are run on the manufacturing line and written into a viewer with only two states. A board where seven tests never ran, including the compressor, the heater and the battery, rendered identically to a clean board in every filter and sort. The log's own summary line said 44 of 44 passed. The correct answer was 36.
Firmware is pushed to units one at a time, by hand, by the person who owned the tooling. A routine act sat in a queue behind an individual, and customer success couldn't answer "is that unit on the current version" without interrupting an engineer to find out.
Three accordions, one per setpoint, fixed fields, nothing addable or removable. The procedure requires all three points. Letting someone delete one produces a certificate that looks complete and is not. Rigidity was the feature.
The approach
Relatively similar for each problem.
Sit with the team that does the thing and have them walk it as they do it now, including the parts they have stopped noticing, because that is where the workarounds live. Get it out of their heads and onto something we can both point at. Then spend most of the time on the edge cases, which is where the decisions turn out to be.
Thermal verification
How Artyc confirms that a shipper holds temperature the way it claims. A unit goes into a chamber and runs through a profile. A hot hold, then a sub-freezing hold, with eleven internal sensors recording throughout. A run is typically eight hours. The question at the end is whether the sensor's readings stayed within the profile's range to call it verified. Dozens of those runs are on record now. The old output of each was a number somebody read off and typed into a document.
The design makes the verdict computed and the evidence permanent. Pass comes off the per-sensor statistics rather than off a person's judgment, and the statistics stay on the page so the next reader can disagree with the verdict without rerunning the test.
Calibration
Calibration is done by a third-party lab. A technician plugs the unit into a reference probe and runs it through three setpoints: frozen, refrigerated, and room temperature. At each setpoint they record the unit's reading against the reference, check whether the gap is within the 3°C limit, and issue a certificate when all three pass. That certificate is the record that a shipper reads true.
The procedure has a fixed shape, so the interface follows it. Three accordions, one per setpoint, each collecting the same fields in the same order. No Add, no Remove. The procedure requires all three points; a form that lets someone delete one lets someone produce an incomplete certificate.
Because the lab is outside the company, the evidence arrives from outside too. Certificates upload against the unit directly, so the reading and the paperwork live in the same object rather than in someone's inbox.
A result that's out of spec always gets written down, and the system says out loud what happens next. Recheck the reading, route the unit for repair. A system that makes a bad result hard to record teaches people to record something else.
Board tests
Five states replace two: pass, warn, fail, blocked, and not applicable, each judged against a versioned bench profile that says which skips are expected on which rig. A skip is either accounted for or it counts against the board. Warns don't decay. They clear by retest or by somebody putting their name to them.
Runs read Passed, Needs attention, or N/A; skips are judged against the bench profile, so expected skips read N/A instead of padding the pass count. The drawer reads the board as subsystems, each carrying its worst result, and a failure comes with what happens next: re-run or escalate.
Firmware
During testing in 2025, a unit downloaded an update, told the app it was up to date, and was still running the old firmware. It took a power cycle to make the claim true.
By 2026 the fragmentation was often commercial. One customer was running several different firmware versions across ten units, and the worry, written down at the time, was that bad results from old firmware "would be an avoidable reason to not land the deal."
The whole design turns on one definition. A shipper is confirmed when data arrives from the new image, not when the shipper says so. A stage can't open until the one before it is confirmed rather than reported.
Where the four meet
The four workflows live together under Tools. The place they actually converge is a single unit. Every shipper has a page holding its firmware version, calibration status, board test result, environment, and group. The page is readable by somebody who has never stood at a bench, which is most of the company.
Binary pass/fail made an incomplete board look identical to a good one. Five states, each judged against a versioned bench profile.
What landed
Work that lived with one person became a record anyone can read. They kept evolving because the platform grew every time the hardware did.
Every firmware release is a test of whether staged rollout holds at fleet scale, about 100 units as of early 2026 and double that when the next product lands later in 2026. Calibration keeps meeting new third-party labs, each with a slightly different process. The question will be how much structure to demand before the friction pushes work back into email. Thermal testing keeps adding test types and what comes back from engineers is a request for granularity. Isolate one sensor, zoom into the ten minutes where something happened, put two runs side by side.
One shipper, every record that has ever happened to it. Readable by somebody who has never stood at a bench.
What I’d do differently
Audit the built thing against the spec inside the first week, every time. An audit seven weeks after ship caught an issue with calibration vocabulary. The spec used "Reliability" to describe performance and calibration results. Reliability is a feeling. Repeatability, on the other hand, has a number attached. A standard deviation limit inside the cycle window, something an auditor can check. The word caused confusion with third parties and QA teams from customers. Nothing about that walkthrough needed to wait seven weeks. It was a quick fix.
Artyc employees can move a unit between databases. For everyone else, the column doesn't exist.
What’s still open
The one I'd chase before any of it is still open against compliance. Whether the person who uploads a calibration certificate can also be the person who approves it. Everything else here is a mechanism. That one decides whether the records the mechanism produces would survive somebody asking.
Note: To respect confidentiality and regulatory compliance, interface images are built against the product's own design system and populated with fixture data. No customer records, colleague names, or partner names appear anywhere in this case.
Engineering platform
Artyc
A staged push any engineer can run. Pick the build, pick the stage, put a name and a reason on it. The next stage opens when data arrives from the new image and not when a shipper says it updated.
A staged push any engineer can run. Pick the build, pick the stage, put a name and a reason on it. The next stage opens when data arrives from the new image and not when a shipper says it updated.
The short version
Artyc makes battery-powered coolers for medical transport that hold temperature for days without plugging into anything. Proving that means chamber runs that bake and freeze a unit for hours, calibration against a reference probe, board tests on the manufacturing line, and firmware going out to units already in the field.
The mechanical, firmware, and software engineers performing this work had workflows running in disparate places, none of which produced anything the rest of the company could easily read.
A platform underneath all of it was missing. Run the thing, record what happened, hand it to somebody else, and still be able to find it a year later. We needed consolidation and tidiness because not having results readily available is as good as not having results at all.
My role was to work across the engineering disciplines that don't always share a vocabulary. A mechanical engineer describes a run by what the chamber did. A firmware engineer describes the same run by what the unit reported. Getting those to be the same record was most of the job.
Firmware became a staged, gated rollout that any engineer can run, not just the one who wrote it. Thermal verdicts are computed from the statistics on the page. Calibration accepts outside lab evidence against the unit itself. Board tests no longer have a two-state verdict, so a skipped test no longer counts as a pass.
Eleven sensors, eight hours. Verdicts compute from per-sensor statistics. Evidence stays visible so a later reviewer does not have to retest to disagree.
The problem
Each team had its own version of the same failure.
Thermal verification generates good data and no way to pass it on. Whoever ran the chamber had the answer and everybody else had to ask them for it.
Calibrations are performed by a third party, so the evidence that a shipper reads true arrives from outside the company. Certificates landed in email with no object to attach them to.
Board tests are run on the manufacturing line and written into a viewer with only two states. A board where seven tests never ran, including the compressor, the heater and the battery, rendered identically to a clean board in every filter and sort. The log's own summary line said 44 of 44 passed. The correct answer was 36.
Firmware is pushed to units one at a time, by hand, by the person who owned the tooling. A routine act sat in a queue behind an individual, and customer success couldn't answer "is that unit on the current version" without interrupting an engineer to find out.
Three accordions, one per setpoint, fixed fields, nothing addable or removable. The procedure requires all three points. Letting someone delete one produces a certificate that looks complete and is not. Rigidity was the feature.
The approach
Relatively similar for each problem.
Sit with the team that does the thing and have them walk it as they do it now, including the parts they have stopped noticing, because that is where the workarounds live. Get it out of their heads and onto something we can both point at. Then spend most of the time on the edge cases, which is where the decisions turn out to be.
Thermal verification
How Artyc confirms that a shipper holds temperature the way it claims. A unit goes into a chamber and runs through a profile. A hot hold, then a sub-freezing hold, with eleven internal sensors recording throughout. A run is typically eight hours. The question at the end is whether the sensor's readings stayed within the profile's range to call it verified. Dozens of those runs are on record now. The old output of each was a number somebody read off and typed into a document.
The design makes the verdict computed and the evidence permanent. Pass comes off the per-sensor statistics rather than off a person's judgment, and the statistics stay on the page so the next reader can disagree with the verdict without rerunning the test.
Calibration
Calibration is done by a third-party lab. A technician plugs the unit into a reference probe and runs it through three setpoints: frozen, refrigerated, and room temperature. At each setpoint they record the unit's reading against the reference, check whether the gap is within the 3°C limit, and issue a certificate when all three pass. That certificate is the record that a shipper reads true.
The procedure has a fixed shape, so the interface follows it. Three accordions, one per setpoint, each collecting the same fields in the same order. No Add, no Remove. The procedure requires all three points; a form that lets someone delete one lets someone produce an incomplete certificate.
Because the lab is outside the company, the evidence arrives from outside too. Certificates upload against the unit directly, so the reading and the paperwork live in the same object rather than in someone's inbox.
A result that's out of spec always gets written down, and the system says out loud what happens next. Recheck the reading, route the unit for repair. A system that makes a bad result hard to record teaches people to record something else.
Board tests
Five states replace two: pass, warn, fail, blocked, and not applicable, each judged against a versioned bench profile that says which skips are expected on which rig. A skip is either accounted for or it counts against the board. Warns don't decay. They clear by retest or by somebody putting their name to them.
Runs read Passed, Needs attention, or N/A; skips are judged against the bench profile, so expected skips read N/A instead of padding the pass count. The drawer reads the board as subsystems, each carrying its worst result, and a failure comes with what happens next: re-run or escalate.
Firmware
During testing in 2025, a unit downloaded an update, told the app it was up to date, and was still running the old firmware. It took a power cycle to make the claim true.
By 2026 the fragmentation was often commercial. One customer was running several different firmware versions across ten units, and the worry, written down at the time, was that bad results from old firmware "would be an avoidable reason to not land the deal."
The whole design turns on one definition. A shipper is confirmed when data arrives from the new image, not when the shipper says so. A stage can't open until the one before it is confirmed rather than reported.
Where the four meet
The four workflows live together under Tools. The place they actually converge is a single unit. Every shipper has a page holding its firmware version, calibration status, board test result, environment, and group. The page is readable by somebody who has never stood at a bench, which is most of the company.
Binary pass/fail made an incomplete board look identical to a good one. Five states, each judged against a versioned bench profile.
What landed
Work that lived with one person became a record anyone can read. They kept evolving because the platform grew every time the hardware did.
Every firmware release is a test of whether staged rollout holds at fleet scale, about 100 units as of early 2026 and double that when the next product lands later in 2026. Calibration keeps meeting new third-party labs, each with a slightly different process. The question will be how much structure to demand before the friction pushes work back into email. Thermal testing keeps adding test types and what comes back from engineers is a request for granularity. Isolate one sensor, zoom into the ten minutes where something happened, put two runs side by side.
One shipper, every record that has ever happened to it. Readable by somebody who has never stood at a bench.
What I’d do differently
Audit the built thing against the spec inside the first week, every time. An audit seven weeks after ship caught an issue with calibration vocabulary. The spec used "Reliability" to describe performance and calibration results. Reliability is a feeling. Repeatability, on the other hand, has a number attached. A standard deviation limit inside the cycle window, something an auditor can check. The word caused confusion with third parties and QA teams from customers. Nothing about that walkthrough needed to wait seven weeks. It was a quick fix.
Artyc employees can move a unit between databases. For everyone else, the column doesn't exist.
What’s still open
The one I'd chase before any of it is still open against compliance. Whether the person who uploads a calibration certificate can also be the person who approves it. Everything else here is a mechanism. That one decides whether the records the mechanism produces would survive somebody asking.
Note: To respect confidentiality and regulatory compliance, interface images are built against the product's own design system and populated with fixture data. No customer records, colleague names, or partner names appear anywhere in this case.