Naweed Kabir
Home CV Thinkaloud (a separate site)

The reasoning behind the result.

I coached people to show AI how to use a computer, and wrote an essay on what it taught me. My thesis asked what a PDF malware classifier bases its verdicts on.

Research2023–2026

Questions I keep asking

  1. What would it take for anyone to hand AI a hard problem and trust what comes back?
  2. How do people learn to show and judge work well enough for an AI to learn from it?
  3. When a model gives an answer, what would let the person relying on it question that answer?
  4. How do we build technology for every kind of mind, including the ones it was never designed around?

What entitled you to say that?

A correct result can hide a missing observation or an unsupported claim. The essay sets out ten habits for catching both, and the question that carries over to anyone working with AI.

Practitioner essayObservations 2022–2025About 19 minutes

Read the essay

Short of time? Start with its ten questions

Looking inside a malware classifier

PDFGuardian compared four models for classifying PDF files and used SHAP to ask what each verdict rested on. A 2026 re-check with the saved models and data took the explanations further.

B.Sc. thesisBRAC University2023

The question
Can a model that flags a malicious PDF also show what it looked at?
The method
Four classifiers compared on a file’s structural features, with SHAP to explain single verdicts and whole models.
Revisited
In 2026, the saved models showed that an explanation’s setting changes its answer.
Why it matters
Someone has to act on the verdict. They need a way to question it before they accept it.
Read the study overview