Skip to content

Hover for colour, click to enlarge.

How I got here

I came for the mind and stayed for the body carrying it

My undergraduate degree was in electronics and communication engineering, which meant sampling theory, reconstruction, and the question of what you are allowed to conclude from an incomplete measurement. At the time I filed that under the dry part of the syllabus.

What I actually cared about was consciousness. I wanted to know how something built out of tissue produces an interior life, and MRI was the first instrument I found that let me look at the thing itself rather than argue about it. So I went toward brain imaging for reasons that were, honestly, more philosophical than technical.

The interest moved. Somewhere between learning how a spin echo works and learning what a Jacobian determinant means, I stopped being fascinated by the mind and started being fascinated by the body carrying it, and by the machinery underneath both. Not only what a method scores, but why it works at all, and what it is quietly assuming while it does.

That is the habit I have not been able to put down. When something works I want to know which part of it is doing the work. When it fails I want to know whether the model was wrong or the measurement was never going to carry the answer. Most of what I have built since is an attempt to make that question answerable instead of arguable.

The field I landed in is medical image computing, and I care about the whole of it rather than one corner. Anatomy and what makes it vary between people. Segmentation, registration, reconstruction, and the geometry and optimisation sitting under all three. How an acquisition decides what a method can possibly recover. What a metric is actually measuring when it agrees with you.

Registration is where that curiosity currently lives, because it states the thing plainly. Every person has a left putamen; no two are the same shape. A smooth invertible map carrying one anatomy onto another has to do exactly the work of that difference, and what the map does is the difference. Atlases, morphometry, longitudinal tracking and label transfer all rest on that one operation. It is quiet infrastructure, and I like that about it.

The sampling theory turned out not to be the dry part. It turned out to be the foundation.

Why a doctorate

I want the time to do this properly

A doctorate is not the only way to work on hard problems, and I thought about that for a while before deciding. What settled it is that the questions I have ended up caring about are the slow kind. They need a control built before the result is believed, an instrument built before the property is claimed, and the freedom to spend a month establishing that something is true rather than shipping the version that looks true. That is what a PhD is for, and I have not found another arrangement that protects it.

Before graduate school I spent two years building medical device software, including a real time engine processing 100 Hz sensor streams and the ISO 13485 calibration and quality control tooling around it. That work taught me something no paper had. The distance between a method that works on curated data and a method somebody can use on a patient is enormous, and it is almost entirely invisible from inside a lab. I would like to spend a decade closing parts of that distance rather than admiring it.

The specific thing I want to get good at is building models of anatomy that are continuous, physically constrained, and honest about their own uncertainty, then establishing where those properties actually matter rather than where they merely sound impressive. My own work has already made that distinction concrete for me: I built the control that showed my headline claim was broader than the evidence, and the narrower claim that survived is worth more than the one I set out to prove.

The reason is concrete rather than aesthetic. Someone reading a longitudinal study deserves to know which part of a measured change is real and which part is the algorithm being confidently wrong. A deformation field that could report where it should not be trusted would change how these methods get used, because the case nobody catches is never the obvious failure. It is the warp that is smooth, plausible and quietly incorrect.

I am not attached to one organ or one modality, and I would rather not become so this early. The problems I find interesting are the ones where the measurement is incomplete and the physics is not optional: sparse views, limited angles, undersampled acquisition, anisotropic sampling. Those are the same inverse problem wearing different clothes, and they show up nearly everywhere in medical imaging. Learning to recognise that shape across domains seems more useful to me than learning one domain very well and nothing else.

What I am learning next

Named plainly, so it is clear where I am heading rather than only where I have been.

Physics constrained modelling

My regularisers are already borrowed physics: kinetic energy, bending energy, a log barrier holding the Jacobian positive. Replacing a generic smoothness prior with the governing equations of the tissue itself is the natural next step, and I would like to learn to do it properly rather than by analogy.

Calibrated uncertainty

My correspondence objective is already a heteroscedastic estimator with a learned precision, but it is calibrated only in the sense that it helps optimisation. Whether a model stated confidence means anything is a question I do not yet know how to answer, and it is the one I most want to.

Reconstruction from incomplete measurements

Sparse view, limited angle, undersampled k-space, anisotropic acquisition. These are the same inverse problem in different clothes, and my results keep pointing toward the regime where sampling is the binding constraint rather than model capacity.

Learning across modalities and scales

I have worked on MRI, CT, chest radiographs and histopathology, and each one taught me that the hard part is rarely the architecture. I want to understand what actually transfers between them and what only appears to.

Education

Where I studied

M.S. in Computer Vision

January 2025 to December 2026
University of Central Florida

GPA 3.889 out of 4.00. Academic advisor Dr. Niels Da Vitoria Lobo. Thesis on continuous diffeomorphic registration with implicit neural velocity fields.

Coursework

  • CAP 5516 Medical Image Computing
  • CAP 5610 Machine Learning
  • EEL 5820 Image Processing
  • CAP 5415 Computer Vision
  • CAP 6419 3D Computer Vision
  • MAP 6197 Mathematical Introduction to Deep Learning
  • CAP 6412 Advanced Computer Vision
  • CAP 5115 Virtual Reality Engineering
  • CAP 6908 Independent Study
  • CAP 6411 Computer Vision Systems

B.E. in Electronics and Communication Engineering

2019 to 2023
Visvesvaraya Technological University

GPA 3.50 out of 4.00. Sampling theory, reconstruction and projection geometry, which I thought at the time was the dry part of the syllabus and which turned out to be the foundation for everything since.

Coursework

  • Digital Signal Processing
  • Digital Image Processing
  • Signals and Systems
  • Data Structures and Algorithms
  • Computer Architecture

Experience

Where I have worked

Teaching Assistant, CAP 4453 Robot Vision

Fall 2026

University of Central Florida, Orlando, Florida

Teaching assistant for Robot Vision under Dr. Niels Da Vitoria Lobo, the undergraduate computer vision course at UCF.

  • Support students through projective geometry, edge and region processing, motion, shading and texture, object detection and recognition, and the convolutional approach to vision.
  • Hold office hours, grade programming assignments, and help students debug their first real vision pipelines, which is usually less about the algorithm than about what their data is actually doing.
  • Explaining convolution and the image gradient to somebody encountering them for the first time has been the fastest way I have found to notice the gaps in my own understanding.
TeachingComputer visionProjective geometryCNNs

Research Assistant, Medical Image Analysis

January 2026 to present

AI MIND Lab, University of Central Florida, Orlando, Florida

First author on QueryPhi, a continuous diffeomorphic registration model currently under review at NeurIPS 2026, and on the anisotropy study that followed it.

  • Designed and implemented QueryPhi end to end: a 5.5M parameter model that decodes a coordinate continuous stationary velocity field with a modulated SIREN over a three level correlation pyramid, integrated by scaling and squaring.
  • Trained fully unsupervised on dual H100 nodes, with MIND-SSC, multi window local cross correlation and a confidence weighted correspondence objective. Segmentation labels are used only for scoring.
  • Reached the highest mean Dice of all ten evaluated methods on IXI atlas registration at 0.7726, and transferred zero shot to OASIS with the best boundary accuracy in that comparison.
  • Built the first quantitative instruments for continuity in neural deformation fields, since every paper in this line asserts continuity and none of them measure it.
  • Wrote a benchmark harness placing thirteen competing methods behind one interface with identical metrics, splits and output schema, so the comparison is genuinely like for like.
  • Automated a 150 run experiment matrix with deterministic job generation, SLURM arrays and resumability, plus four parallel Optuna studies of 160 trials across 16 concurrent workers.
PyTorchSIRENDiffeomorphic registrationSLURMH100Optuna

Research Assistant, Vision Language Evaluation

August 2025 to January 2026

Center for Research in Computer Vision, UCF, Orlando, Florida

Co-author on VISTA, an interaction aware benchmark for spatio-temporal understanding in vision language models, at the CVPR 2026 PVUW workshop.

  • Curated roughly 12,000 video and query pairs into one taxonomy that decomposes a clip into entities, their actions, and the relations between them.
  • Built the harness that runs 11 state of the art vision language models end to end, normalises their free form generative output, and scores it automatically.
  • Ran the diagnostic breakdown that surfaced spatio-temporal patterns which aggregate accuracy had averaged away.
Vision language modelsBenchmark designEvaluation

Computer Vision Engineer

2023 to 2024

Equidor Medtech LLP, Bengaluru, India

Built and shipped EquiCOG, a real time binocular eye tracking system for oculomotor and vestibular diagnosis, as regulated medical device software.

  • Processed 100 Hz synchronised sensor streams from embedded hardware in real time, with live filtering, artifact and dropout rejection, and derived motion metrics.
  • Shipped ISO 13485 compliant calibration and quality control tooling whose automated checks flag a bad acquisition while the patient is still in the room rather than a week later, cutting device setup time by 25 percent.
  • Built the hardware accelerated FFmpeg pipeline for synchronised video and signal logging, and authored the traceable design and test documentation the standard requires.
  • Working next to clinicians taught me how large, and how invisible, the gap is between a method that works on curated data and one somebody can use on a patient.
OpenCVReal timePyQtFFmpegISO 13485

Research Intern, Signal and Data Analysis

2022 to 2023

Cyclops Medtech Pvt. Ltd., Bengaluru, India

Analysed large scale clinical ocular movement recordings and built the pipelines that replaced manual spreadsheet workflows.

  • Built cleaning, feature extraction and batch processing pipelines in NumPy and Pandas over device time series.
  • Refined the vestibular diagnostic algorithms against real patient recordings rather than synthetic test signals.
NumPyPandasSignal processingTime series

Toolkit

What I work with

Grouped by what it is for rather than by how often it appears on a job posting.

Medical image analysis

  • Deformable and diffeomorphic registration
  • Stationary velocity fields, scaling and squaring
  • LDDMM and B-spline parameterisations
  • Implicit neural representations, SIREN, FiLM modulation
  • Segmentation: U-Net, Attention U-Net, 3D U-Net, SwinUNETR, DynUNet
  • Foundation model adaptation: SAM, LoRA
  • Similarity metrics: LNCC, MIND-SSC, NGF, mutual information
  • FreeSurfer, ANTs, NiftyReg, SimpleITK, MedPy, nibabel
  • NIfTI and DICOM handling

Modelling and mathematics

  • Differential geometry of deformations
  • Jacobian determinants, log Jacobian dispersion
  • Variational regularisation: diffusion, bending, kinetic energy, log barrier
  • Cost volume and optical flow methods, PWC-Net and RAFT style
  • Fourier analysis, sampling and reconstruction theory
  • Wavelets, Haar and multiresolution decomposition
  • Hessian spectrum and loss landscape analysis
  • Heteroscedastic and Bayesian uncertainty modelling

Deep learning

  • PyTorch
  • MONAI
  • scikit-learn
  • Optuna hyperparameter search
  • Weights and Biases
  • TensorBoard
  • Multi GPU and distributed training
  • Mixed precision, gradient checkpointing
  • LoRA and parameter efficient fine tuning
  • Vision transformers, Swin, nnFormer, CoTr

Statistics and validation

  • Subject level bootstrap confidence intervals
  • Paired Wilcoxon signed rank tests
  • Holm correction over pre-specified families
  • Friedman and Nemenyi critical difference ranking
  • Wilson intervals for rates
  • Cochran-Armitage trend and logistic interaction tests
  • Cross validation and unit of analysis reasoning
  • Ablation and attribution design

Languages and tooling

  • Python
  • MATLAB
  • C and C++ fundamentals
  • TypeScript and JavaScript
  • Bash and shell automation
  • LaTeX
  • Git and GitHub
  • NumPy, SciPy, Pandas
  • Matplotlib, Seaborn, Plotly
  • OpenCV, scikit-image, Pillow
  • Adobe Lightroom, Photoshop, Premiere Pro
  • FFmpeg

Systems and infrastructure

  • Linux
  • HPC and SLURM job arrays
  • GPU accounting and fair share budgeting
  • Docker
  • Dual H100 and A100 nodes
  • Reproducible experiment tracking
  • Regression checks on scientific results
  • Next.js, React and Tailwind
  • Real time 100 Hz embedded sensor pipelines
  • PyQt, FFmpeg, ISO 13485 documentation

Outside the lab

Long integrations, faint signal

I photograph the night sky. It is the same problem as my day job wearing a different hat: pull a faint signal out of noise using long integrations, careful calibration, and a great deal of patience in post.

The Milky Way core over the treeline at Kissimmee Prairie Preserve State Park

The core rising over Kissimmee Prairie Preserve State Park, Florida, the only International Dark Sky Park in the state and about the darkest sky within driving distance. Hover to play.

Camera
Nikon D5600
Focal length
17 mm
Aperture
f/2.8
Exposure
30 sec
ISO
1000
Mode
Manual
Metering
Centre weighted average
Exposure bias
0 step

Astrophotography taught me things about my own field that no course did. A faint target rewards patience rather than force. You stack many exposures because signal adds coherently while noise adds in quadrature, which is the same reason a claim about registration accuracy needs a hundred paired cases rather than one striking example. You shoot calibration frames because your instrument has a systematic character you must measure and remove, which is the same reason a comparison needs a matched control rather than a published number.

And you learn to be careful with a beautiful result. A Milky Way frame can be pushed until it looks spectacular and means nothing, in exactly the way a deformation field can be smoothed until it looks plausible and has stopped describing the patient. Knowing where that line sits, in both cases, is mostly a matter of having crossed it before.

It is also simply what I do when I want to think about something other than a loss curve. Driving out to Kissimmee Prairie Preserve, the only International Dark Sky Park in Florida, waiting for the core to clear the horizon, and being reminded that the interesting problems are not all inside a GPU.

The sky is where the patience goes, but it is not the only thing I shoot. A fair amount gets taken on whatever is in my pocket, usually on nights and roads that were not planned. More of that will end up here.

Stacking is averaging with better manners

Signal adds coherently across frames while noise adds in quadrature. That single fact is why a hundred short exposures beat one long one, and it is the same reason a claim about registration accuracy rests on 115 paired cases rather than one striking example.

Processed in

Adobe LightroomPhotoshopPremiere Pro
Milky WayStackingCalibration frames

Also true

Away from the desk

Where the thinking happens when nothing is being graded.

Where I go when I am not working

Games are the one place I get genuinely unsupervised. Not a break from thinking so much as a different room to think in, with no reviewers and no one to convince.

What I keep going back to is anything that trusts me to be somewhere and work it out. Sekiro, which is only ever about paying attention. Ghost of Tsushima, for what it does with quiet. The Last of Us. Prototype, which I loved for less defensible reasons. An open world I can get lost in for an evening is close to my favourite thing there is. Marvel Rivals, Valorant and League happen too, though mostly those are an excuse to be around friends.

Mostly seinen

Mostly seinen, some shonen when it earns it. What I am after is the same thing I am after in a game or a paper, which is a piece of work willing to sit with a difficult idea rather than resolve it early.

I will follow almost anything that is genuinely introspective, that takes its own premise seriously and does not flinch from where it leads. It is the closest thing I have to a reliable test of whether something was made carefully.

I travel alone, on purpose

Mountains, weather that most people would call bad, and nights. Somewhere far enough out that the sky is the largest thing in view and there is nothing to do but look at it.

The moon, and past it whatever is further out. It is the same pull that has me driving four hours to photograph a part of the galaxy I will never get closer to, and it is where most of my actual thinking gets done.