Reflections

An infinity mirror is a simple trick done carefully: a fully mirrored back panel, a half-mirrored front panel, and a ring of LEDs in the gap between them. The light bounces back and forth, losing a little each pass, and you get a tunnel receding into the wall. Mine is six inches square: two 150 mm acrylic panels, 32 addressable LEDs, a wood frame cut from 2×4 lumber, about 44 grams of printed PETG holding the electronics.

This is the second one I’ve built. The first was a twelve-inch hexagonal piece, and it’s where all of this started. This six-inch tile is where that object became a product.

What makes this one different is what happens when you break the mirror on purpose. Send me a logo or a line drawing, and a tool I wrote laser-cuts that shape clean through the mirror tile, then 3D-prints a clear PETG plug that drops into the cutout flush with the front face. A thin base layer behind the mirror bridges the cut fragments and holds them: the kintsugi glue. Lit, the tunnel runs to infinity everywhere the mirror survives, and your shape glows where it doesn’t. The flaw is the feature. That’s the kintsugi idea, borrowed honestly: the repair is where the light gets in.

The design part

I spent eleven weeks in an independent study turning this object into a near-sellable product, running three lanes at once: the product, the business, and the customers. The thesis was that I was designing a system, not a gadget, and the brand had to say that too.

The identity I locked is literal about it. The company is Layered Logic: layers of print, of mirror, of process, of data. The mirrors are its first product line, Reflections. Five adjectives govern every decision (ambient, considered, quiet, crafted, living), with an indigo anchor and a type system where italic is the default voice and upright is the gear-shift into the technical register. The tagline is a motif, not a fixed string: light that layers for Reflections, with the noun swapped per product line. The rule underneath all of it: the tech disappears. This is an object for people who want presence in a room, not a spec sheet on a wall.

The decisions that mattered most weren’t visual. Writing a right-to-repair philosophy before designing the electronics meant the controller became a replaceable module, the power supply a commodity barrel-jack brick, and the whole unit openable with one Phillips screwdriver. The customer repair guide covers all seventeen catalogued failure modes. The warranty commits to parts for seven years. None of that is marketing. It’s constraints I gave myself early, and the hardware is shaped by them.

The engineering part

The firmware is a from-scratch build on the ESP32-C3: six modules, written to principles locked at the start. It works with no smart-home connection, the app is the pattern editor, and it updates over Wi-Fi with a reflash backstop.

Hardening it was the best engineering of the quarter. Four measured passes on the live control path: a 30 ms broadcast debounce, fanout moved to a dedicated FreeRTOS task so one slow client can’t stall everyone, a per-IP rate limit, then characterization under reconnect storms, injected latency, and injected packet loss. First-response latency dropped from 155 ms to 84 ms. The final soak ran 1,476 iterations with zero errors. Every step was build, OTA to the real mirror, and measure. Nothing counted until it ran on silicon.

Device pairing is authenticated with HMAC-SHA256 across the firmware, the Android app, and the webapp, backed by a pure-C crypto reference with 21 known-answer tests so every signer agrees byte for byte. And one honest failure stays in the record: I tried three ways to stream OTA progress to the app, each one broke the install, and I parked it cleanly rather than ship something flaky. Knowing when to stop is part of the process.

The fabrication side got the same treatment as the firmware: build the machine that builds the thing. One customer SVG fans out into four fabrication files (laser cut paths for the acrylic, a two-layer printable plug), with corners pre-rounded to what the nozzle can actually produce and the laser’s kerf compensated. Parametric generators emit the packaging dielines for any box size, validated ten-for-ten against reference cut files. Zero per-order CAD. There’s a live version where you can upload your own art and preview the real, manufacturable result, not a fantasy of it.

The company part

I ran this as a real LLC, not a class exercise. The Washington certificate of formation was filed before the quarter started; EIN, operating agreement, sales tax registration, and a compliance checklist with re-check dates followed. The BOM is costed at four volume tiers: materials run about $28 a unit at one-off quantities and fall toward $16 at a thousand, with the custom controller PCB landing near a dollar a board at production volume. The uncomfortable finding: labor beats materials at every tier, so the honest retail anchor sits around $65 for the basic unit, and the real opening is custom work north of $200, where a repairable design object has no competitor.

What the quarter actually taught me is where my edges are. Engineering verified on real hardware, systems thinking, automation, documentation: those came easily. Cold outreach and user interviews did not; the contact list sat stale for weeks while the firmware got four polishing passes it didn’t strictly need. I know which of those is the harder problem now, and it isn’t the firmware.

Where it stands

The quarter ended the way it was designed to: a complete, sellable system. Service blueprint mapped end to end, supply chain sourced with backups, policies drafted, custom pipeline live. It is one switch flip from a store.