I build software that
talks to machines.

A controller from the 1980s has a few kilobytes of buffer and no way to say how full it is. You get one warning — XOFF — and it arrives late.

Too fast and the machine silently drops characters while cutting steel. Too slow and a 40-minute job becomes four hours. The engine treats this as a control loop: it speeds up while the machine drains and backs off the instant XOFF arrives.

The engineering log records the debugging done on live machines — including the fix that created the next problem.

There is nothing Windows-specific inside the engine: the same tree compiles and passes its tests on Linux, Windows and macOS on every commit.

Android has no serial port. Reaching RS-232 means USB OTG, an FTDI chip and a JNI bridge into a Java library.

Receiving worked immediately. Sending did not — and none of the five causes raised an error: a zeroed timeout parked the reader thread so writes vanished silently; a poller fought for the connection a thousand times a second with no lock; and a flow-control request was rejected over a name mismatch, then carried on as if it had worked.

I confirmed that last one by reading the bytecode of the library itself.

Six FreeRTOS tasks and lock-free ring buffers separating the serial line from the card and the web server — so a slow SD write could never interrupt the byte stream feeding the machine.

The tablet replaced the device, but the firmware stays public: it is the same problem solved on bare metal.

When a controller throws an alarm at 2 AM the answer exists — in a 400-page PDF nobody can search, or in the head of the technician who is asleep.

A wrong answer here injures someone, so an explicit safety policy runs before replying. A document only enters the index after human review. And there are AI cost ceilings per request, per day and per month, so a retry loop cannot burn the budget overnight.

118 automated tests run on every commit.