Profiling says the kernel is memory-bound, not compute-bound
Two hours of profiling beat a week of guessing. Global reads are uncoalesced; shared-memory tiling is next, then a Triton comparison for honesty.
- p50
- 7.1 ms
- baseline
- 8.7 ms
Field notes
What I'm working on
Ongoing public activity: build sessions, benchmarks, architecture decisions, business experiments and the lessons from things that failed.
SPH boundaries stopped leaking
Boundary stiffness tuning fixed particle escape at large timesteps without dropping below the frame budget.
- frame
- 14.8 ms
- particles
- 1.2M
Every commerce number should be derivable from the event log
If a total cannot be rebuilt from raw events, it is an opinion, not a number. NFN now replays deterministically.
A failed run is still data
The first batched-verification attempt was slower than the naive path. Publishing that saved the next three days.
Unit economics before acquisition
Shuya Life margin model rebuilt from real logistics costs rather than assumed ones. Two SKUs removed.
16 Augarchitecture
Systems → Products → People → Profit → Freedom
The whiteboard version of why any of this exists: systems that compound, products that ship, results that are measurable.