What actually went wrong, what we changed, and what it cost. These are write-ups of decisions made on real production hardware, not summaries of somebody else's news.
Before television, my mother copied passages out of newspapers into a notebook. My father recorded cassettes before leaving for work. Both were technology used for education. This is what changed and what did not.
A big screen plus voice opens more product space than embedded hardware, iterates in days instead of months, and parents accept it where they resist a tablet. Also, why we cut the read-aloud test.
Fur-and-motor companion pets fail in the hand, not on the spec sheet. Three layers have to land together, and below a certain score the product is worse than not existing. A note on closing a road you cannot walk.
Fifty children in a classroom are loud enough. Fifty AI devices talking at once is a different problem, and it is not one you invent at a desk. What screenless meant, why the knowledge base is a gate, and what mute-all actually required.
Two small displays were not a styling decision. Silent on a table, a voice device looks like any speaker. What eyes solve is the silent state, and why offline expressions matter more than online ones.
A general model answers with whatever was most common in its training data. In medical devices, chemical reagents or insurance, that is a liability. What a bounded knowledge base changes, and which businesses it fits.
The first enclosure was modelled on an AirPods case. Outdoors you had to hold it to your ear. Why amplifier tuning could not save it, and what we gave up to make the product audible.
Bluetooth was ruled out by the chip. 2.5mm turned out to have no stock anywhere. Mono came from a requirement nobody writes down. A selection trail for anyone building guided-tour or companion hardware.
Embedding the SIM on a 4G device looked like the responsible choice. It blocked operator partners, made repairs a teardown, and raised cost in the exact markets we were targeting.
We added a dedicated noise-reduction IC to let users interrupt a talking toy. Here is why distance, volume, voiceprint and DSP each fail differently, and what we shipped instead.
A project that failed because users had to read out an 8-digit student ID. What broke, why a screenless device made it worse, and the one question we now ask at every interaction step.
