Writing

Technical notes on machine learning and the engineering decisions behind finished work.

ML Under the Hood

• Backpropagation Without Computational Graphs — A step-by-step explanation of backpropagation using equations, the chain rule, and gradient accumulation.

Engineering Notes

• Why asynchronous model loading was three times slower — contention, polling, copies, and parameter resolution turned concurrency into overhead.
• How to report temporal-model throughput without ambiguous FPS — prediction cadence, consumed frames, and latency are different measurements.
• What a 10× pipeline speedup does and does not prove — end-to-end throughput improved; component attribution requires another experiment.