The product was itself a validation article — built to worst-case size, mass and electrical load, so the receiving system was tested against a known article.

Production tooling would not arrive in time for the first builds, so I split the architecture by validation need and held the interfaces between configurations fixed.

Strain goes as t/L², force as t³/L³. Halving the thickness cuts strain 2× and force 8× — safe, and it feels loose. Lengthen the beam instead.

The screw holds 1449 N. The polymer under its head gives out at 121 N, and the handbook torque puts five times that through it. The answer is a flange head, a compression limiter, or a torque spec written from the bearing calculation rather than the fastener table.

Insert moulding gives the best pull-out and torque-out and the worst serviceability. Anchorage does not have to mean a knurled threaded insert: on the components I worked on it came from undercuts and geometry designed so the resin locks the insert in every direction it could otherwise leave. The wall around it still has to be thick enough that shrink lands as compression rather than hoop tension, and the recurring failure is resin flash onto a face that has to stay clean — held out only by a tight shutoff.

Two-shot injection moulding is the same idea without a separate part: the second resin is moulded onto the first, so the bond and the mechanical interlock are designed into the geometry. It removes an assembly operation and a tolerance loop, at the cost of a more complex tool and a material pair that has to be compatible.

I temperature-soaked the interfaces and measured heat transfer into the finger with a thermisthesiometer, a skin-analogue contact probe, feeding the CEM43 thermal dose equation — so Industrial Design, safety and engineering argued in the same units.

Two dynamic components — here only Component A and Component B — had to hold a tight axial alignment, positioned independently by an X-locator and a Y-locator, each with its own long chain. Because they interacted dynamically, the misalignment was perceptible to a user.

I analysed it worst-case rather than RSS: this interface had to align under every permitted condition, and a small tail of misaligned units is still a user-facing defect.

The analysis drove changes to architecture, dimensions, interfaces and datum strategy. Industrial Design reviewed the resulting mechanical behaviour and accepted the revised design.

A separate set of production-intent consumer products, not the validation article above. Across them I built the methods that turned mechanical feel into measurable requirements.

Preload. Everything between the press surface and the switch sits inside a tolerance band, and where that stack lands decides how much force is already on the dome. Too much and F₁ arrives early with a shallower snap; too little and there is lost motion.

Press location. Users do not press the centre. The levers are the support and guide scheme, press-surface stiffness, where the switch sits relative to the thumb, and how much tilt the guides allow — none of them the switch itself.

Detent feel is the same problem about another axis, as torque against rotation: peak resisting torque, how deep the assisting region runs, seat pitch, and consistency seat to seat and in both directions.

Assembly demanded precision the available process could not hold. Several supplier fixture concepts failed and it declared a no-build condition — one operation driving takt time, and no poka-yoke to stop a bad unit leaving the station looking correct.

No working alternative, and the production component was already released — so I solved it as a process problem, using properties the components already had.

The schedule. The build was on site at an overseas vendor and the team was flying out for it.

I also owned production definition for critical components: GD&T drawings, tolerance analysis, datum strategy, FAI requirements, SPC callouts, MSOP definition and build-to-print release.

I reviewed supplier FAI reports and resolved out-of-spec dimensions with them, deciding case by case whether the part, the process or the drawing was wrong.

Drawings, FAI reports, dimensions, tolerances and product-specific features are not shown.

The tooth counts are not arbitrary. The first layout paired a 22-tooth pinion with the 82-tooth ring. Both even, so a pinion tooth only ever meets half the ring teeth — the same half forever — repeating any form error at a fixed angular phase and concentrating its wear there.

Dropping one tooth fixes it, at a 4.8% change in ratio — upward, which the torque budget wanted anyway — and half a module of motor-mount offset to hold centre distance.

The two radial features are deliberately not equal: the ball groove runs 0.1 mm of clearance, the bearings 0.2 mm. The tighter one picks up first, so the race locates the plate and the bearings only retain it axially and catch an overload. Three bearings each fixing X and Y would over-constrain it.

Specifying that as a relationship rather than two numbers is what makes it survive printing: both features drift with the same process bias, so the groove stays tighter than the bearings wherever they land.

The pinion rides on the moving plate and the ring gear is fixed, so the plate’s radial position is the centre distance of the mesh — and centre distance sets backlash. At a 20° pressure angle, 0.1 mm of float is 0.073 mm of backlash variation.

A pinion inside an internal ring gear gives continuous rotation and a compact package, but it has clearance between the teeth by definition, and every reversal spends that clearance before any motion reaches the turret.

Three candidates are open: preloading the existing pinion as a split gear, or replacing the drive with a cycloidal or harmonic reducer. Nothing here is decided.

The split gear is the option I have specified furthest, because it is the one I can build and test soonest.

The pinion is already herringbone, so the apex where the two opposed helices meet is a natural parting line. Split there and each half is a single helix of opposite hand: pushing them apart along the shaft wedges each against the flank it faces, and their axial thrusts still cancel through the hub.

An anti-backlash gear only holds zero backlash up to its preload. That rules the approach out for power transmission, but the yaw drive moves a light head against about 2.4 N at the pinion, so the preload sits comfortably above the working load.

What settles it. How much backlash the current base actually has. Whether the head can be repackaged to free the ring bore, which is the only thing keeping cycloidal off the table. And whether the split gear’s preload drag fits the stepper’s thermal margin.

A stepper turns a lead screw; motor and screw pivot on a trunnion at the base so the screw can swing, and the travelling end drives a linkage that swings the head about a trunnion at the top of the gusset.

The screw is single-start with a fine lead so that it self-locks: the axis holds elevation unpowered and the barrel does not drop on power loss. A worm drive is the only other option that does that, at several times the mass — and every gram here rides on the yaw axis.

The payload preloads its own drive. The range never approaches vertical, so the gravity moment stays one-signed and holds the screw against the same nut flank permanently. The drive never crosses its own backlash, so the pitch axis needs no anti-backlash hardware.

Keeping the payload CG a few degrees below the bore axis preserves that. Worst-case gravity moment is 87% of peak over this range, against 57% if the axis reached vertical — a flatter load, bought by not asking for elevation the application does not need.

Why a drop-off cam. Energy goes into the spring over most of a revolution and comes out in milliseconds, so the motor only has to cock. Rate of fire is set by how fast the cam turns.

Why the barrel moves. A conventional layout needs a feed mechanism, a valve trip and a cocking stroke. Making the barrel the moving element collapses them into one: its position is the gate, its travel is the trigger, and the cam motor is the only actuator.

The CO₂ valve is off the shelf; valve design was outside the scope of this project. A future version may use a custom valve so the discharge can be tuned rather than worked around.