In-house role
Deterministic learning engine
Swapped an unstable AI runtime on two partner learning platforms for a deterministic engine: the same output, with no inference cost at render time.
The story
The problem
Two partner learning platforms rendered their lessons through an AI model at run time, and the output was unstable.
What I did
- Replaced the runtime model with a seeded, deterministic block transform.
- Proved the new engine output-equivalent to the old one before switching over.
- Put the lesson blocks behind one registry, and forked the second platform from the first in a single day.
The result
- Zero inference cost at render time.
- Every module renders the same way, every time.
- The lesson I keep: AI is not the answer to a formatting problem.