Hardware-aware Pseudo-Softmax: a research note

This note organizes a hardware-aware view of approximating Softmax: the target is not only model accuracy, but an implementation that can be measured, verified, and mapped to efficient hardware.

The relevant questions are about approximation error, memory, parallelism, RTL constraints, and the impact of replacing an expensive operation inside a larger model.

A research note for organizing the intuition behind approximating Softmax when the goal is efficient and verifiable hardware.

Raul Pacheco Rodriguez

Anti-leakage evaluation for reproducible medical AI

In medical AI, a strong number can hide a weak experimental question. This note frames anti-leakage evaluation as a prerequisite for reproducible and trustworthy systems.

The practical direction is to document data manifests, split data by the right real-world constraints, and report failure modes instead of only the best metric.

Initial notes on temporal shortcuts, information leakage, and more honest evaluation protocols for medical AI systems.

Why I am separating this research blog from FP32

This English version keeps the core idea of the original note: FP32 can remain a broader editorial space, while this site can host work that is closer to research, papers, systems, and professional updates.

The distinction matters because not every idea belongs in the same publication context. Some notes are exploratory, some are technical logs, and some are part of a longer academic trajectory.

A short note on why this site should host research-adjacent writing while FP32 remains an independent editorial lab.