ABOUT ME

The people I interviewed about wheelchairs kept telling me about pain, and pain was the thing nobody had an answer for. Years later the same shape kept appearing: autoimmune conditions, gut trouble, chronic pain, sensory sensitivity, and certain neuropsychiatric diagnoses turning up together in the same people and being treated by five different specialists who never speak to each other.

A long essay written with Luca Estinto, treating these as a diseasome — a set of conditions linked not by the organ they present in but by the pathway underneath them — and arguing that sensory processing is a good deal of what that pathway is. A nervous system which never habituates, and keeps reporting a signal at full strength, produces consequences that cascade through the immune and digestive systems by way of stress. The piece pulls together imaging work on altered connectivity, the habituation literature, glutamate as a shared route to overexcitability, microbiome composition differences, and the enteric nervous system’s direct hand in immunoglobulin secretion.

A cat of mine was dying because his colon was paralysed, and the two options put in front of me were surgery and euthanasia. He was too heavy and too unwell to survive the anesthetic, so in practice that was one option. There is a third, and it is not in the literature: it circulates as folklore between practitioners, never reaches a journal, and so whether it is offered to your animal depends almost entirely on who you happen to be standing in front of that day.

An owner-reported case study, written up in enough detail to take to a vet and ask. Recurrent idiopathic megacolon: the colon dilates and the nerves that should drive it stop producing any useful push. Enemas, manual deobstipation, laxatives and dietary modification all failed to prevent recurrence. The regime that held was neostigmine, an acetylcholinesterase inhibitor borrowed from equine practice, injected when palpation finds accumulation — it leaves acetylcholine in the junction longer, so the cholinergic signal telling the muscle to contract actually lands. Alongside it: daily lactulose to keep the stool soft, and an anti-NGF injection every six weeks for spinal spondylosis.

Sequencing matters more than the drug. Neostigmine is given only after the colon has been emptied under anesthetic and mechanical obstruction has been definitively excluded, because a drug that makes a bowel contract harder is dangerous if that bowel is genuinely blocked. Duration to treatment is the other variable: medical management succeeds in about two thirds of cats presenting under six months of symptoms, and in under six percent of those presenting later. Maintenance is injections every few weeks, indefinitely.

People who use mobility aids told me they were choosing every day between being stoned past the point of finishing a sentence and being lucid and in agony. I wanted to know whether anything avoids that choice.

An argument with references, written to be taken to a doctor or a funder. Memantine blocks the same channel as ketamine, gently enough to take daily for years, at six cents a tablet. The trial evidence is thin and mixed: it fails in old nerve injury and shows promise where pain is central and ongoing. The post lays out the mechanism, every trial including the negative ones, the commercial history that ended the research, and the trial that would settle it.

A mouse cannot report a symptom, but it is talking constantly in a register we cannot hear. Two things follow from listening properly. You get some understanding of what these animals are actually saying to each other. And you get far more out of each experiment while doing far less to the animals in it: a microphone over a cage takes readings continuously, at night, in the dark, from an animal nobody has touched, picked up, restrained or moved to a testing room — which is both a kinder life and a cleaner measurement, since a frightened mouse is not the animal you meant to study. More information out of every cohort, less disturbance per animal, and a clearer picture of what a compound does before it goes anywhere near a person.

Computational bioacoustics at Vium — unsupervised classification of ultrasonic vocalizations in quasi-real time, which is also how I dipped my toes into audio processing. This was early: several years before MUPET and DeepSqueak made unsupervised clustering the standard way to do this, and as far as I know the first to run it in quasi-real time on cages being monitored continuously rather than on a corpus after the fact. Letting the categories fall out of the recordings instead of deciding in advance what to listen for is the only way to hear a call nobody has named yet. The calls turned out to carry the animals’ libido and their stress, so I could discover and then implement a way of reading those off the audio directly. It was deployed in Vium’s preclinical studies, where it could tell you whether a compound shifts libido or stress before the drug ever reaches a human trial. Two write-ups came out of it, A New Female–Female Mouse Vocalization Discovered via Unlabeled Machine Learning and On the Detection and Prevention of Aggression in Lab Mice via Quasi-Real Time Analysis.

A multi-company validation in 2025 put three compounds that had passed conventional safety pharmacology under continuous non-invasive monitoring in their own cages and found signals anyway, some persisting for days after dosing. Quietly watching animals who are left alone turns out to see things that handling them on a schedule does not.

Neuroprosthetics have two major issues. At the tissue end, putting an electrode into a brain is an injury: microglia arrive, astrocytes proliferate, and a glial scar closes around it, pushing the neurons you wanted away from the thing that is listening — so the signal thins and gets noisier across the months in which a person is coming to depend on it. At the other end, any single neuron tells you almost nothing; the intention to move is legible only across a whole population of them. I came at this from the software side and assumed both were software problems. My first thought was that a decoder which recalibrated fast enough could just track the signal as it drifted — that convergence was the thing to fix. I worked on that for a while before accepting you cannot infer your way out of an electrode with fewer and fewer neurons left to hear.

The software half was convergence analysis of the decoders then in common use: how fast a Kalman-style filter settles onto a usable mapping from neural activity to intended movement, how far that mapping degrades as the underlying signal drifts, and how much of the loss you can claw back by letting the decoder keep adapting while somebody is using it. The number that matters to a user is not peak accuracy on a good day. It is how long recalibration takes every morning, and whether the thing is still working by the afternoon.

The hardware half was optical recording: instead of pushing metal into cortex and waiting for the scar, you get the neurons to express a fluorescent indicator that brightens when they fire, and read them with light. It buys cell-type specificity and thousands of cells at once. It costs you depth — a millimeter or so even with two-photon, in an organ several centimeters thick — and temporal resolution, because calcium rises and falls far more slowly than a spike does, and it requires getting a gene into somebody, which is a much larger request than a connector.

Most of the answer has been materials. The stiffness mismatch is absurd — silicon around 180 GPa against a brain of a few kPa — so the field went soft: ultraflexible polymer threads, mesh electronics injected through a syringe, carbon fibers seven microns across, substrates stiff enough to go in and then soften once wet. In mice it works beautifully. One open-mesh design tracked the same neurons for thirteen months, most of the animals’ adult lives. Coatings do the rest: PEDOT to drop impedance, neural adhesion proteins, dexamethasone eluted to quiet the response. Then the uncomfortable part. Across fourteen BrainGate participants and up to seven and a half years, the arrays held spiking on about a third of their electrodes and declined only about seven percent — and when arrays are taken out and examined, the leading failure is the silicon eroding and the metal coming away, not the tissue closing in. Softer probes may be the right fix for the wrong failure. “Scar-free” was a mouse result in 2017 and has been quoted a long way past its evidence. Meanwhile the thing these arrays turned out to decode best was not a limb but speech, which in 2014 I would not have guessed at all.

For someone with celiac disease or a severe food allergy, every meal they did not cook themselves is an act of trust in a stranger, and the cost of that trust being misplaced runs from a ruined week to a funeral. The idea was to hand the judgment back to the person eating, in something small enough to live on a keyring.

I tried it twice. The first attempt was optical: a one-dimensional array of infrared and visible lasers with an avalanche photodiode, looking for gluten’s absorption signature, which would have read food about a centimeter deep without touching it. It drowned in noise. Gluten’s signal is faint against everything else going on in a protein-rich food, and protein-rich food is exactly where you need to look.

So I switched from physics to chemistry: G12 antibodies, which bind gliadin, read out by a colorimetric assay — a toothpick you poke into the food, or a strip that works like litmus paper. G12 is less sensitive than A1 and much cheaper, and cheap is what matters when the thing has to work at every meal. That is where it stopped, and the wall was the antibodies: short of bulk synthesis, a per-meal consumable costs more than the people who need it can spend.

In March 2014 I found GlutenTox, TellSpec and 6SensorLabs already at it — the good kind of disappointment. 6SensorLabs became Nima, which shipped in 2017, missed gluten at the legal threshold more than a fifth of the time, died in 2020, and came back in 2026 claiming 10 ppm, by its own measurement. My wall got routed around rather than climbed: the expensive part was the antibody, and the field swapped it for aptamers, synthetic DNA that folds around a target for a fraction of the cost. One 2024 design reads gliadin off gold nanoparticles with a phone camera. The part underneath is still unsolved: a negative on one bite does not clear the plate, and sampling error swamps measurement error.

The idea was to help autistic people read social communication: if the signals everyone else is supposedly picking up on are real, a machine ought to be able to surface them.

A Google Glass app that recovers a person’s pulse from the video feed alone, no contact and no cuff, and reads the changes in it while they talk. Joint with Kartik Talwar and Spencer Hewett.

I would not build this now. It takes a reading off someone’s body without their knowledge and gives them no way to refuse — the person being measured is the one person in the room with no say in it. The need was real; meeting it this way moves the cost onto somebody who never agreed to carry it.

My grandfather lost his mobility, so I volunteered at a spinal cord injury rehabilitation unit and asked the people there: What would improve your quality of life? Three answers kept coming back: getting between bed and chair, being at eye level in a conversation, and ground rougher than the floor the chair was built for. Separately I went looking at what actually kills people with spinal cord injuries. Mostly infection, and an open pressure sore is one of the ways infection gets in. They are not the biggest killer — that is pneumonia; they are the biggest killer mechanical engineering can do anything about.

Everything was modular, and that was the point: not a new wheelchair, but parts that bolt onto the one somebody already owns. A chair gets fitted to a body over years, and asking someone to give that up in order to gain a feature is not a trade most people will take.

• five robotic pressure-injury relief mechanisms, shifting the occupant’s load without them having to do anything — depending on someone to remember it, which many cannot, fails on exactly the days they are unwell
• a transfer mechanism, for getting between bed and chair
• a seat that rises, for talking at eye level
• a drive and tires for ragged ground

The year after, I mentored Ada Rosa on the same problem from the other end: mobility assistance for people with ALS and spinal cord injuries, driven by gaze, for users whose remaining reliable movement is their eyes — here she is showing off the eye control.

I expected to hear about getting around. What I kept hearing about was pain — constant, and largely unaddressed by anyone they had seen.

It was not a thing I could answer with mechanical engineering. What I could do was make sure the mechanisms were not locked up: I wrote up all the mechanical engineering and soft robotics I’d invented as a nonprovisional patent application, Robotic Mobility Assistive Wheelchairs, and then decided not to file it, and they went out under a Creative Commons license instead. A patent would have meant the people I built them for waiting on somebody else to license it first.

Then I went to work on understanding and treating chronic pain, which I am still doing.

Both halves have since arrived, about a decade on. Medicare began covering powered seat elevation in 2023 — on the grounds of transfers and reach, not of meeting people at eye level, which is what the users I spoke to actually said. And automated pressure redistribution that asks nothing of the occupant is now a pilot study. I should say plainly that all of this, mine included, measures interface pressure, which is a proxy; nobody has shown it prevents an injury.

Blindness in robots. A machine that stops dead the moment its camera fails is a machine you cannot trust around people, and sensors fail constantly.

So I programmed the PR2 to keep working with its other sensors switched fully off — no vision, no external feedback of any kind — and to learn, under those conditions, to place objects into the holes of the corresponding shape. All it had was its own past motor position commands and the finger gripper sensors, so the robot had to adapt to the task from its own sense of where its body had been.

Our team also smoothed joint movement of the Willow Garage Personal Robot 2 and improved load equalization — sharing the torque between its two arms so it could hold something bulky whose mass sat off to one side without one arm taking the whole of it. That is most of what carrying things for somebody actually is: a bag of shopping is heavier at one end, a box is packed unevenly, and a person handing it over expects the thing taking it to adjust.

What is recoverable once a sense is gone?

The Braille work, the following spring. Here, a robot reaching for a hole it cannot see and finding it by touch; there, a language meant to be read by touch, worked out from fragments by a machine that cannot see either. Both are asking what is recoverable when the channel everyone assumes you have is missing.

To know where you are, you need a map. To build a map, you need to know where you are. So how does anything ever get off the ground?

You do not solve either one first. You estimate both at once and let them correct each other. Take two frames from slightly different positions, triangulate the handful of features visible in both, and you have a crude map — which means you have to move before you can know anything at all, because depth comes from parallax. After that it alternates: given the map, work out where the camera must be to be seeing what it sees; given the poses, add new points and refine the old ones. The tracker running underneath this was PTAM, whose trick was to split those two jobs across threads — follow the features every frame, rebuild the map in the background whenever there is time to spare.

Error accumulates the whole way, because every new position is measured against a map built out of the previous ones. You get it back by recognizing somewhere you have already been and closing the loop. And there is one thing a single camera can never tell you: scale. A large room far away and a small room close up produce the same picture. That number had to come from the sonar altimeter and the inertial unit.

All of it ran off-board, on a laptop, over WiFi. It held together where there was texture and steady light, and fell apart where there was not: repetitive ground, changing sun, and wind, which moves the aircraft between one frame and the next.

I would not take this on now. Navigating terrain nobody has surveyed is not a neutral capability — there is very little daylight between solving it for a toy quadcopter and solving it for something armed. I was treating it as a puzzle about maps and motion, not thinking of the consequences.

My first robotics project, and the motive was homesickness — I missed my dog at university, so I built something that would trail after me.

One told it to follow you by standing in front of it. Two problems follow from there: knowing who to keep following, and keeping up with them.

The first answer was color — lock onto the histogram of whatever the person was wearing and chase that — which works right up until they turn away, walk under a different lamp, or take their coat off. So it identified them by gait instead, from the way they walked rather than from their face or their clothes.

For keeping up I took what looked like the cheap way out. Six legs are statically stable: an alternating tripod keeps three feet on the ground with the body over its own support at every instant, so in principle the thing never has to balance at all. In practice it fell over constantly. Static stability is a promise about geometry, not about the world — it assumes the feet land where you expect, that the ground holds when they do, and that you are moving slowly enough for momentum not to count. A slope, a patch of gravel, a foot dropping into a dip a centimeter deeper than the controller believed, and the support polygon is not where it is supposed to be. Extra legs buy margin, not immunity.

The way out was not more legs, even though I personally find that cuter. Boston Dynamics went the other way with BigDog — four legs, which cannot fall back on geometry at all and so have to catch themselves, force-controlled limbs choosing where to put a foot in the next fraction of a second. That is the machine everybody remembers getting kicked sideways on ice and skittering back upright, and it is the approach that generalized. Quadrupeds now learn the recovery in simulation and cross ground I could never have got a hexapod over. Following a person around has a name now too — robot person following — with benchmarks, and re-identification that keeps learning your target as they change.

What if, rather than planning motion across a floor, you add flight to the problem? Three dimensions, no ground to stand on, and no option to stop and think.

A superconductor carries current with no resistance: nothing lost, nothing turned to heat, and magnets far stronger than their power draw. Every one we have needs liquid helium, so the chase is for one that works warmer — and you cannot run that chase by building candidates one at a time.

Extending GMU’s simulation after transferring there, advised by Dimitrios Papaconstantopoulos, using electronic-structure calculation by the augmented plane wave method to predict the superconducting transition temperature of intermetallic compounds. This became one of my two undergraduate theses, Predicting Superconductivity Transition Temperature via the Augmented Plane Wave Model. Alongside it, algorithms of AI chess players.

The approach paid off, in other people’s hands. H₃S was calculated to superconduct near 200 K and then measured at 203 K; LaH₁₀ was predicted in 2017 and found at 250 K two years later. Materials predicted on a computer before anyone made them, which is the outcome the whole approach was betting on. They need megabar pressures, so nobody is wiring a city with them yet.

A multiphysics simulation is several different physics solved over one domain at once, and it is expensive twice over: the run itself, and the human work of stitching solvers together without introducing errors. Both of those are structural costs, and the structure is already sitting there in the mathematics.

The proposal was to write these models in a dependent type language, because the mathematics is already the right shape for it. A classical field is a section of a bundle over the domain, and a gauge field is a connection on one; a bundle is a fibration. A dependent type is also a fibration — a family indexed over a base, which categorically is a map whose fibers are the types. The two are the same construction wearing different clothes, so a physics model can be carried in a type rather than in a comment, and the compiler can then see what the physicist knows.

The payoff I was after was run time, not just correctness. Once the structure is in the type, whole categories of runtime work become statically unnecessary — dimension and index checks, guards on couplings that cannot occur, dispatch that can be specialized away. That is not speculative: eliminating array bound checks through dependent types measurably speeds up ordinary numerical programs. A multiphysics solver has far more structure to exploit than a bubble sort does. Making direction of dependence part of the type — which physics feeds which — is what pushes it from ordinary dependent types toward the directed, cocartesian kind.

Other people were heading the same way, and one of my own co-authors on Toward Directed Collapsibility is among them: Nicole Sanderson now writes on topology and neural circuits. Persistent homology has reached the tumour microenvironment too, picking out rare T-cell states under checkpoint blockade. The immunotherapy daydream had other people in it all along.

Both halves have since been developed, independently of me. Directed type theory was rebuilt on synthetic simplicial foundations, where the objects under study are exactly cocartesian fibrations, and the proofs are going into a proof assistant. On the applied side, Decapodes composes systems of PDEs diagrammatically: you assemble a multiphysics simulation out of component models instead of editing a monolith, and the composite is a first-class object rather than a hand-merged file. They report a drastic drop in time-to-first-simulation, though they say plainly that the claim is qualitative and unmeasured, which is the honest state of it.

Space is cool, and a satellite that folds is cooler: you launch something small and it opens into something large. But a deployed satellite is a geometric mesh sitting in the ionosphere, which is a plasma, and that plasma is part of the electrical environment the structure has to work in. So the question is what a given geometry does to the fields around it.

I was given access to a hollow aluminum polyhedron, a vacuum pump and a 2.45 GHz RF source: evacuate it, backfill with argon, and strike a discharge inside. The cavity’s geometry picks out which modes can live there, and the discharge organizes itself along the resulting standing wave. The point was the mode structure: which field patterns a given polyhedron will and will not permit. A benchtop analogue for mode selection rather than a faithful orbit — argon is the conventional stand-in for low-earth-orbit plasma, but it matches neither the ion mass nor the temperature, and a microwave discharge sits orders of magnitude above ionospheric density.

People do this deliberately now and call them plasma metasurfaces: discharges arranged so their geometry decides which electromagnetic modes get through, and reconfigured by changing the discharge rather than the hardware. I was making a very crude one and calling it a side project.

A memristor is a resistor that remembers where you last left it. That sounds like a small thing and it is not: a component that stores a value and computes with it in the same place means a device built from them does not have to send its data somewhere else to be processed. Put a memristor on a flexible surface and that device can be worn: potentially as a non-invasive medical device.

Chua predicted the device in 1971; HP said they had built one in 2008; I was depositing titanium dioxide films in 2012, when the field was still mostly promise. What I did was make them and measure them — bench work, not simulation, a research assistantship in the materials science lab of the Chemistry and Physics Department at Mary Baldwin College — pushing each one until it switched, then bending it and doing it again to find where the behavior fell apart. It is also where I became interested in physical examples of nonlinear systems.

Years afterward, a group wired live rat neurons in Padova to silicon neurons in Zurich through memristors in Southampton, with the memristors holding the synaptic weights between them. The devices were Pt/TiO_x/Pt — the same stack I had been making by hand. I had nothing to do with it, and it is still my favorite thing to have happened to a material I once held.

This started because I read Braille and kept noticing that the signs around me were wrong. The one that decided it: a door labeled electrical room in print and safety exit in Braille. Whoever installed it could not read what they were mounting, and nobody downstream of them could catch it. So I wanted translation to be something a builder could get right at the point of installing a sign, rather than something they had to take on faith.

The mistakes are easy to make, because Grade 2 Braille is not a cipher you can look up letter by letter — it is contracted, with about a hundred and eighty contractions, and the same six dots can be a letter, a whole word, or a fragment glued to its neighbors depending on where in the word it falls and what sits beside it.

Which makes it a very good test case for a problem much older than Braille: you have a text in a script you only partly understand, and a translation of some of it, and you want the rest. The method does not care that the script is Braille. It builds probabilistic dictionaries from whatever parallel fragments exist and lets the grammar fall out of the alignment — the same shape of problem as an incompletely deciphered ancient script, where a handful of confident readings and a great deal of unread text is exactly the situation you are in. Braille simply has the advantage of a known answer to check against. It grew out of my automated computational semantics research at GMU and became my other undergraduate thesis, Contextual Machine Learning through the Analysis and Chunking of Partially Translated Grade 2 Braille.

The grammar was the problem, and the standards body agreed. When Unified English Braille replaced the older code, nine contractions were deliberately deleted — in BANA's own words, to enable accurate automatic translation and reduce the exceptions to the rules. They simplified the language rather than the model. It is still not solved: frontier language models today refuse, hallucinate, or emit malformed Braille, and a small fine-tuned model beats all of them.

I kept meeting working doctors who were further from the current literature than they wanted to be. Not for any want of caring: seeing patients all day and reading a stack of new papers every week are not two things one person can do. The volume is what defeats them, so the volume is what to go after.

Automated computational semantics research at George Mason, aimed at getting a machine to read a paper and return something faithful and short enough that a clinician between appointments could decide whether the whole thing was worth an evening.

This is the research CAMEL grew out of, pointed the other way round. There, meaning recovered from too little text; here, meaning kept intact while most of the text goes away.