Before becoming a doctoral researcher, I worked in law enforcement as a cybercrime investigator. That experience still shapes the kinds of AI systems I want to build and, just as importantly, the kinds of claims I am willing to trust.
Evidence changes how you think about AI
Investigation work leaves little patience for fluent but weakly grounded conclusions. A claim matters only insofar as the evidence behind it can be inspected, challenged, and connected back to a concrete chain of events. In research, it is easy to be impressed by a strong benchmark number or a polished output. In operational work, that is not enough.
That instinct transfers naturally into AI. One reason I find NIST AI RMF 1.0 useful is that it does not reduce trustworthiness to accuracy alone; it treats reliability, explainability and interpretability, privacy, and accountability as parts of the same design problem. When the cost of a mistake is high, traceability matters. You need to know where a claim came from, what uncertainty remains, and how a human can verify it.
Adversarial thinking is useful beyond security
Cybercrime investigation also teaches you to expect adaptation. Fraud tactics change. Attackers respond to defenses. Surface-level success can hide deeper fragility. A system that looks reliable under static assumptions may fail quickly once the environment becomes strategic.
That mindset is valuable well beyond traditional security. It encourages questions like:
- What happens when the data distribution shifts?
- Where can the model leak information?
- What assumptions are we quietly making about reliability?
- How easy is it to inspect a failure?
Why this matters for LLM systems
Large language models are powerful, but in high-stakes contexts they should be treated as components in a system, not as the system itself. Retrieval, evaluation, and evidence presentation are not optional extras. They are what make the system usable without breaking the chain of evidence.
In that sense, I am sympathetic to Cynthia Rudin’s argument in Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. In serious settings, the goal should not simply be to attach better post hoc explanations to opaque systems. The better question is whether the system can be designed to remain inspectable enough for the task in the first place.
This matters in domains like fraud analysis, speech intelligence, and security investigation, where the difference between a useful output and a misleading one often comes down to whether the reasoning remains grounded, reviewable, and proportionate to the stakes. The OECD AI Principles reinforce the same point by linking trustworthy AI to transparency, robustness, accountability, and meaningful human oversight.
A better framing
The transition from investigation work to AI research changed my technical interests, but it did not change the standard I care about. I still want systems that help people reason better under uncertainty rather than merely automate impressive-looking outputs.
That is why I am drawn to speech intelligence, retrieval-augmented systems, trustworthy AI, and AI security. They all sit near the same core question: how do we make machine intelligence useful without breaking the chain of evidence?
Closing note
I do not think of my background in cybercrime investigation as something separate from AI research. I think of it as the reason I care so much about building AI systems that are evidence-aware, secure, and genuinely trustworthy. The tools have changed, but the standard has not: in high-stakes settings, useful intelligence should be grounded enough to inspect, cautious enough to question, and practical enough to support human judgment rather than replace it.