Open to summer 2027 internships

Circuit boards, and the machine learning that runs on them.

Second-year MEng Biomedical Engineering student at UCL. I design hardware, train the models on top of it, and spend most of my time checking whether the numbers are real.

See the work Read the preprint

UCL MEng Biomedical Engineering First-class first year Sole-author preprint with DOI

383 / 710

distinct images found in a published clinical benchmark of 710 records

0 / 24,110

false accepts across impostor faces in the drone's re-identification pipeline

55 / 59

labelled bounces detected on real table-tennis footage, plus 39 of 39 net crossings

< 40 dB

the noise regime where learned EIT reconstruction beats the classical method

Selected work

Six projects, each with a number I can defend

Every figure below comes from a script that regenerates it. Where a result is simulation only, or from a single venue, it says so.

Preprint · Zenodo DOI

PallorHb: anaemia screening from conjunctival images

A study on the public CP-AnemiC dataset that turned into a dataset-integrity audit. The 710 published records are 383 distinct photographs; duplicates crossed collection sites and carried conflicting haemoglobin labels on identical pixels.

Grouping folds by site, the standard safeguard, still left two-fifths of a +0.176 AUROC inflation in place.
Corrected baseline AUROC 0.706 (95% CI 0.653 to 0.757); 68 manuscript claims verified automatically.
Code and the duplicate detector released.

View code Read the preprint →

Python · scikit-learn · NumPy · SciPy · bootstrap CIs · DeLong · Bland–Altman · grouped CV

Figure from the preprint: the same conjunctiva photograph filed under two identifiers at two hospitals with two haemoglobin values, and the near-zero residual after a 3-pixel shift

One photograph, two identifiers, two hospitals, two haemoglobin values. Images: CP-AnemiC dataset (Asare, Mendeley Data), CC BY 4.0.

Render of the PURSUIT-1 flight-controller PCB, an X-shaped four-arm board with motor rings at each corner

Hardware · Firmware · Vision

PURSUIT-1: a person-following quadcopter whose airframe is the PCB

The 4-layer flight-controller board is generated from a Python circuit description, so the netlist and layout have one source of truth. All checks pass and fabrication files are released.

ESP32-S3 flight firmware: sensor fusion, cascaded PID, motor mixing and failsafes, validated on the bench.
Off-board detection, tracking and face re-identification on ONNX Runtime with CoreML, chosen over PyTorch to keep the deployment at 40 MB.
0 false accepts in 24,110 impostor faces; 369 tests.

KiCad · SKiDL · Freerouting · ngspice · C++ · Python · ONNX Runtime · OpenCV · PlatformIO

Instrument · Inverse problem

EIT-16CH: an electrical impedance tomography instrument

A 16-electrode, 4-layer measurement board with 104 components, DRC clean and released for fabrication, alongside a simulation study of how images should be reconstructed from it.

Learned and classical (Gauss-Newton) reconstruction compared across noise levels, in simulation.
The learned method wins below 40 dB SNR; above it the classical method holds its own.

KiCad · ESP32-S3 · PlatformIO · PyTorch · NumPy · pyEIT

Grid of EIT reconstructions: ground truth, classical Gauss-Newton and the learned method at clean, 30 dB and 20 dB noise

Two frames from a real match with the detected ball track drawn on: bounces marked in green, net crossings in magenta, racket hits in red

Detected track and events on real match footage. Frames: OpenTTGames (OSAI), CC BY-NC-SA 4.0.

Computer vision · Validated on real footage

TT-Scout: table tennis match analytics from video

A classical ball detector, tracker and velocity-kink event detector for bounces and net crossings, with no trained model. Player identity by shirt colour, no faces.

On public footage (OpenTTGames, two clips, one venue): ball on 94 to 96% of labelled frames, 55/59 labelled bounces at 2 to 4 cm median error.
39/39 net crossings; 4/4 winners match the umpire's scoreboard.

Python · OpenCV · NumPy · Matplotlib

Reinforcement learning · Uncertainty

RL Car: uncertainty-aware autonomous racing

A racer that estimates its own epistemic uncertainty and slows when a sensor is blinded. The useful result was a negative one.

Uncertainty has to be estimated on a forward model, not the policy: the policy-space ensemble scored AUROC 0.127, anti-correlated, on a blinded sensor.
A Navier–Stokes solver written from scratch for the body aerodynamics, validated against published cylinder-flow benchmarks.

Python · NumPy · SciPy · bootstrap ridge regression · random Fourier features · FreeCAD · gmsh · CalculiX

Render of the RL car: lofted blue body shell over a printed chassis, differential-drive rear wheels and a pivoting front beam, front three-quarter view

150 by 90 by 46 mm, body lofted through seventeen cross-sections; differential drive at the rear, front wheels on a pivoting beam. Rendered from the OpenSCAD model.

Embedded · Modelling

Airlift photobioreactor controller

A 2-layer ESP32-S3 controller for a 1 L algae culture, with a growth model fitted to decide the light schedule.

Optimised schedule predicts +21% biomass at equal energy, in simulation.
Board is DRC clean with fabrication files built.

KiCad · ngspice · CadQuery · Python · PlatformIO

Publication

One preprint, sole author

Duplicate leakage in a public conjunctival pallor benchmark, and an honest baseline for image-based anaemia screening

Htet, P. H. (2026). Preprint, Zenodo. doi:10.5281/zenodo.22782147

Shows that a published clinical image dataset contains far fewer distinct images than records, that the duplicates carry conflicting labels, and that the usual site-grouped cross-validation does not remove the resulting inflation. Reports a corrected baseline and releases the detector.

Open on Zenodo Code and detector

About

Where the work comes from

MEng Biomedical Engineering, UCL

2025 to 2029. First-class result in year one. A-Levels: A* Mathematics, Further Mathematics, Physics, Chemistry; A Biology.

Research

Oral cancer detection with deep learning, with Dr J. W. Asare, University of Energy and Natural Resources, Ghana. 2026 to present.

Six months on the wards

Healthcare assistant at Grand Mandalay Hospital, Myanmar, January to June 2025. This is what turned my interest towards machine learning for clinical care.

Skills

Tools I use every week

Languages

Python, C, C++, MATLAB, SQL

Evaluation

Bootstrap confidence intervals, DeLong test, calibration (Hosmer–Lemeshow, ECE, Brier), Bland–Altman agreement, decision curves, grouped cross-validation, leakage control

Machine learning

scikit-learn, PyTorch; epistemic uncertainty via ensembles and random Fourier features, out-of-distribution detection, reinforcement learning, gradient boosting

Vision

CNN, ViT, DETR, YOLO; OpenCV, ONNX Runtime, CoreML; detection, tracking, face recognition

Electronics

2- and 4-layer PCB layout, analog front ends, power; KiCad, SKiDL, LTspice, ngspice; SMD soldering; ESP32-S3, RP2040

Simulation and CAD

MuJoCo, domain randomisation; Fusion 360, SolidWorks, CadQuery, gmsh, CalculiX; 3D printing

Looking for a summer 2027 internship

Medical imaging, biosignals or embedded systems, ideally somewhere the evaluation has to satisfy a regulator as well as a reviewer. Available full-time from mid-June to late September 2027.

Email me Download CV