The original cleaning system relied on suction from the ingredient pump alone. Any small leak or line imperfection meant the cleaning cycle just failed with no indication. In a commercial deployment, an undetected failed clean is a compliance risk and a service call waiting to happen. The system had no way to confirm cleaning had actually occurred.

I designed a nozzle cap assembly (inspired by the espresso portafilter already in every barista's hands) that seals the nozzle and creates a closed reservoir for cleaning fluid. A four-phase automated cycle runs nightly: air purge, water fill, concentrated clean, and rinse.

The hardest part was the dosing. A cleaner pump adds concentrate from an external jug while a PWM-controlled water solenoid dilutes it simultaneously. Getting the ratio right required balancing solenoid pressure against pump speed. Too much water pressure and the concentrate never reaches the correct dilution. We validated the final ratio across five concentrations using pH strip testing and visual line inspection.

I coordinated with firmware and software engineers on the sequencing logic, solenoid modulation, and the tracking system that logs every clean cycle across all deployed machines. The cap went through several iterations. The espresso handle was added after operators kept complaining about the original design. Early handle versions tore threads out of the plastic under repeated torque, which I solved by switching from stainless inserts (which also stripped) to a purpose-made press-in threaded insert.

Switching to the internal CIP system also unlocked a format change for the cleaner concentrate from small 4oz bottles to a 1-gallon jug. The combination of optimized dosing and bulk format reduced cleaner cost per machine per year by 82%. The new system also gave the team visibility into cleaning compliance across every deployed unit, so we could catch machines that weren't being cleaned before they caused field failures.

The constraint was durometer. Too soft and the heart shifts during transit, risking damage from movement. Too stiff and contact pressure itself risks tissue damage. The margin is narrow and there's no standardized reference to work from. The geometry and compliance requirements are specific to a cardiac preservation device that had never been built before.

I taught myself silicone casting from the ground up, including mold design, mixing ratios, and demolding techniques, because the project demanded a material and process no one on the team had worked with. Starting with DragonSkin 20 and working up through various platinum-cure silicones to Shore A 60, I cast roughly ten prototype slings, each adjusting durometer, wall profile, and cradle geometry.

Low-durometer slings tore during demolding or deformed so much under cardiac weight that the heart shifted freely, defeating the purpose entirely. High-durometer versions were structurally sound but acted like a trampoline: too stiff to conform, they bounced the heart rather than cradling it. The sweet spot required enough compliance to distribute contact pressure evenly without allowing lateral movement.

I validated each iteration using published average, upper, and lower cardiac mass data, loading the sling with representative weights and physically carrying the assembled device to observe movement under realistic transit conditions. The final geometry and durometer held securely across the full weight range without applying damaging contact pressure. I handled full mechanical integration of the sling into the preservation device housing.

The device entered clinical trials. The sling design was accepted as part of the final device specification, solving a restraint problem that had no existing precedent in cardiac preservation.

Operators interact with the container caps more than any other part of the machine, loading and swapping ingredients throughout service. The original design required a screwdriver, three countersunk screws into plastic, and an o-ring seated in a groove. Assembly was slow, error-prone, and frequently done wrong. Sealing was inconsistent because any manufacturing variance caused leaks. Switching to injection molding made things worse: molded parts warped more than machined ones, and the internal geometry made cleaning impossible, blocking NSF approval entirely.

For the custom container cap, I designed for the interaction first: tool-free, no ambiguity about orientation or assembly order. A single injection-molded cap with a press-fit ring that retains a spring-loaded ball valve. The ball seats by default and opens only when depressed by a mating feature inside the dock. Every surface is smooth and accessible for cleaning. The sealing mechanism is the geometry itself: if the part molds correctly, it seals correctly.

The generic cap was harder. It needed to accommodate a much wider range of bottle geometries while staying just as intuitive. I kept it fastener-free using threaded features and added deliberate flex cutouts in the plastic walls, compliance zones that absorb bottle diameter variance without cracking. The caps hold large bottles inverted, so they also have to support significant weight when drawers are opened and closed without tipping or breaking seal. The tradeoff was balancing enough flex for tolerance absorption against enough rigidity to hold a full bottle securely under dynamic loads.

Both caps are fully injection-molded, fastener-free, and hand-disassemblable in seconds. New operators needed no training on the cap system. They passed NSF review, eliminated leak and assembly failures, and expanded bottle compatibility beyond the original design. Customer complaints about the cap system went to zero.

The original sensor used conductive probes inside the ingredient dock to detect liquid by bridging two metal contacts. With water-based ingredients it worked. With syrups, sugars, or dairy, buildup accumulated on the probes within days. Because the probes sat outside the nightly cleaning path, they were never sanitized. Once fouled, they stopped working, and they fouled fast.

We tried every non-contact sensing method available: break-beam, ultrasonic, optical. Same problem every time. Every ingredient eventually coats the interior tubing walls, making visual and optical detection unreliable regardless of the sensing method.

The solution was a custom thermal runout sensor. One RTD element acts as a heater; a second monitors the temperature response downstream. Liquid flowing through the tube draws heat away at a measurable rate, so the sensor detects flow without ever contacting the ingredient. Doesn't matter what the liquid is.

The first attempt used off-the-shelf thermal sensors with copper tape backing, wrapped against the stainless tube with electrical tape. It worked on the bench but was messy, impossible to repeat consistently across units, and the thermal contact varied with wrap pressure. Every sensor performed slightly differently.

That inconsistency motivated building a custom potting jig. We potted the sensing elements directly against the tube using a hot-melt compound, achieving uniform thermal contact without fasteners or snap features. Moving the sensor to just after the ingredient pump placed it inside the cleaning path, so it gets sanitized every night. Off-the-shelf thermal flow sensors ran $50+ per unit — ours cost a fraction of that.

Sensing reliability improved by 95%. The sensor works identically across water, syrups, dairy, and every other ingredient in the machine. Removing the probes eliminated the fouling failure mode entirely and cleaned up a hard-to-service area of the machine. Reliable low-ingredient detection now reduces machine downtime between drink orders.