● saheed faremi · digital profile · v0.0.1

Researcher · Engineer · PhD-track

Saheed Faremi.

Researcher of the brain. Engineer of the systems people rely on.

Read more Or jump to the research

01 — About

A researcher who ships.

Portrait

forthcoming

Portrait — pending

I work at the intersection of cognitive neuroscience and production engineering.

By day I build software people rely on — multi-asset fintech (Curnance), HR for African mid-market employers (Etihuku), healthcare data infrastructure (HIS Core, predict-dx), and learning systems (Skills Hub, Moodle). The through-line is that infrastructure for under-served users — geographically, economically, or computationally — needs the same engineering rigour as infrastructure for everyone else, and tends to need it more.

My doctoral research turns from the systems people use to the people themselves — specifically, what the brain looks like when it's running. EEG microstates are quasi-stable scalp topographies that segment continuous EEG into a discrete temporal alphabet; I'm working on whether deep generative models — variational autoencoders and Gaussian-mixture VAEs — can learn a microstate segmentation that's more stable across sessions and more behaviourally predictive than classical clustering. DRAFT — verify framing

In 2022 I represented Eswatini at the UNESCO India-Africa Hackathon at Gautam Buddha University in Uttar Pradesh. Team Geeks_on_Fire — five countries, five people — won problem statement AGRI12: an AI-assisted voice contact centre that lets farmers without smartphones report issues by phone and receive guidance back in their language. Gold medals and a ₹3 lakh team prize.

Based in Eswatini. Travel for research. open to collaboration

Infrastructure for under-served users needs the same engineering rigour as infrastructure for everyone else — and tends to need it more.

— Working principle

02 — Research

EEG microstates with deep generative models.

EEG microstates are quasi-stable scalp topographies — typically four to seven canonical classes — that segment continuous EEG signal into a discrete temporal alphabet. The classical approach uses modified k-means clustering over the global field power maxima. It works, but it depends on hard choices (the number of states, the reference electrode, the band-pass) and the resulting segmentation can be brittle across sessions.

This project asks whether a learned latent geometry — via a variational autoencoder — produces a microstate alphabet that is more interpretable, more stable across sessions, and more predictive of behaviour than the classical pipeline. DRAFT — confirm with supervisor

Approach

VAE. Single Gaussian latent prior; learn a continuous embedding of topography frames; segment by latent-space clustering or by direct decoder reconstruction error.
GMM-VAE. Gaussian-mixture latent prior — one component per microstate class — so the segmentation falls out of the latent prior structure rather than a post-hoc clustering step.
Architecture search. Sweep over latent dim, regularisation, and decoder choices; compare reconstruction-vs-segmentation tradeoff curves across architectures.

diagram forthcoming

VAE Vanilla variational autoencoder — single Gaussian latent prior.

diagram forthcoming

VAE-GMM VAE with a Gaussian-mixture posterior at the latent level.

diagram forthcoming

GMM-VAE Mixture-of-Gaussians latent prior — class structure native to the prior.

Diagrams export from ~/Documents/*.drawio as SVG and drop into src/lib/assets/; pass via src.