About · Tom Humplik

Molecules.
Batteries.
Wargames.

The path here was not straight. The way I work has been.

Before any of this was about AI, I spent more than a decade on problems that, from the outside, had nothing to do with one another. I studied how a water molecule moves through a pore barely wider than the molecule itself. I worked on batteries whose important physics happens inside materials you cannot see into. Then I left the lab and started building models of military systems, where the parts were aircraft, supplies, infrastructure, and people instead of atoms.

The domains changed. The job underneath them did not: take something complicated, work out what actually matters, build a representation of it, and test whether that representation survives contact with reality.

NOWNational Reconnaissance Office2021 —Director, AI/ML rapid prototyping

The part I keep short.

I direct a rapid-prototyping software team at the National Reconnaissance Office, building AI-enabled analytical systems for intelligence work that runs well past the scale a human analyst can cover. That is about as specific as this page is going to get about the current job, on purpose.

What I can say about the shape of it: the job is less about any individual prototype than about direction. What the team takes on. What an AI capability has to prove before anyone is allowed to depend on it. Which prototypes have earned the right to become something the mission runs on, and which ones were worth building precisely so we could stop.

The scale and the domain are new. The core problem is not: decide what matters, work out what the data can actually support, and build systems that help people reason over complexity without manufacturing certainty.

What I can talk about freely is everything that came before it — and that turns out to be the more useful half of the story, because it is where the way I work got set.

01MIT2008 — 2014PhD, Mechanical Engineering · Device Research Lab

Small enough to disappear.

I arrived at MIT as a mechanical engineer and spent six years working at a scale where mechanical intuition stops being much help. My doctoral research was on water transport through sub-nanometer zeolite pores — channels roughly 5.5 ångströms across, wide enough to pass a water molecule and narrow enough to turn away a hydrated salt ion. The question behind it was whether materials engineered at that scale could make desalination fundamentally cheaper.

In practice that meant synthesizing materials, building the experiments that could measure them, altering surface chemistry a defect at a time, and setting what we measured against what molecular simulations said should happen. The interesting part was almost always the gap between the two.

The dissertation was Investigating Transport Through Sub-Nanometer Zeolite Pores. Years later those experiments were still being cited by people explaining why supposedly extraordinary nanoporous membranes behave so ordinarily once someone builds one.

What stayed with meA model earns its keep by being wrong in a way you can learn from.

02Sandia National Laboratories2014 — 2016Postdoctoral associate

When the system fights back.

After MIT I moved to Albuquerque and traded membranes for batteries. Larger scale, same shape of problem: everything that decides how a battery behaves happens where you cannot watch it — ions working through a porous electrode, materials swelling and contracting, electrical paths quietly failing as the cell cycles.

So we built ways to make the invisible measurable. One project used electron-probe microanalysis to track electrolyte through thermally activated battery electrodes, turning a spatial chemical map into transport properties like permeability and tortuosity. Another turned out to be mechanical rather than chemical: cycling degraded the conductivity of the polymer-and-carbon binder holding a lithium-ion electrode together by 45 to 75 percent, which is a physical mechanism for a performance loss that had been easy to blame on something else.

It was the kind of problem I still like most. A real system, imperfect observations of it, a model of what we thought was happening inside, and an experiment capable of proving us wrong.

What stayed with meIf you cannot observe the thing you care about, design a way to observe its consequences.

03CNA2016 — 2021Research Analyst → Research Data Scientist

From physics to decisions.

CNA was the sharpest turn in my career. I left the laboratory for national-security analysis, and the systems stopped being pores and electrodes and started being platforms, networks, logistics, adversaries, and the people making decisions inside all of it.

It was less of a career change than it looked. The job was still deciding what mattered enough to represent, what could safely be abstracted away, what the data could actually support, and where a model was handing us confidence it had not earned.

One of the public examples is AGILE 17, a wargame built for the Joint Staff’s logistics directorate: how do you move people and materiel around the world during simultaneous crises, while an adversary works to break the network? You cannot reproduce that. You can build a smaller version of it. Routes become a network, resources become pieces, demand becomes something you can measure, decisions change the state of the board — and a room full of people can finally handle a problem that otherwise only exists in slides.

Over five years the work moved steadily toward data science. The instruments became code, data, and statistical models rather than microscopes and material samples. The discipline did not move at all.

What stayed with meThe best model is rarely the most complete one. It is the simplest one that still preserves the decision you have to make.

THE UNITS CHANGED

Ångströms kilometers

The questions didn’t.

Different domains. Same discipline.

Looking backward, the jumps make more sense than they did while I was making them. At MIT I learned to reduce a physical system to the mechanisms that mattered. At Sandia I learned to build experiments around things I could not observe directly. At CNA I learned that both ideas hold up when the system is made of organizations and decisions instead of atoms.

  1. What is actually happening?
  2. What can we measure?
  3. What are we assuming?
  4. Can we build something that tells us whether we’re right?

That is still roughly how I start. The difference now is how fast the last question gets answered: the thing I build to find out is usually running on my phone within a week, and my kids are a merciless test group. The projects on this site come out of the same habit — a football argument, a Formula 1 dataset, a lost tooth, a supply network, a camera in a three-year-old’s hands.

Generative AI looks like another sharp turn, and it has not felt like one. The central problem is the one I have had every time: decide what the model should represent, work out what evidence it can actually support, find the places it fails, and build the machinery around it that lets someone trust the result. A language model is a representation of a system, fitted to data, confident well past the edge of what it knows. I spent a decade learning how to work with things like that before anyone was calling it AI.

See what that produced

SELECTED RESEARCH

More than a thousand citations across the whole list, nearly all of them for work I did long before I ever shipped an app. The full list is on Google Scholar .