A result is only the beginning.
A useful machine-learning result needs enough context for someone else to reproduce it: the model revision, recipe, runtime, hardware, and workload. An accuracy number on its own leaves too many questions unanswered.
Making complex
things work.
I build tools for reproducible machine learning, work on model compression, and write software that knows when to hand control back.
A little about me
I’m John. I care about the distance between a result that works once and a system you can rely on—recipes someone else can rerun, compression that holds up on real hardware, automation that stops instead of guessing. Away from the screen, I ride my red CB350RS, usually toward higher ground.
Things I’m working on
lazybrick
ML infrastructureMaking model compression reproducible. Declarative recipes, explicit execution contracts, and evidence you can inspect.
Life outside
Riding
My red CB350RS and a winding road.
Travel & mountains
New places, unfamiliar trails, higher ground.
Cooking
Making something good, away from a screen.
Gym
Time to move, lift, and put in the work.
Music
A soundtrack for the everyday moments.
Notebook
Smaller doesn’t automatically mean faster.
Fewer bits can save memory. Whether they save time depends on the kernels, the hardware, the decoding overhead, and the workload. I’m interested in where a compression idea survives contact with the system that has to run it.
Knowing when to stop is a feature.
An automation system needs a clear boundary around what it knows. If an action may have happened but there’s no confirmation, blindly trying again can make things worse. Explicit states, verification, and a way to hand control back are part of the design.