The work that is not part of my research line, kept here rather than on the front page: machine learning, embedded hardware, security, and the industry roles I held before graduate school. It is the background that makes the robot fleets in my research possible.
The stack I build with: hover or tap a component
The surface a non expert actually touches: a canvas to draw a motion on, a library of what they have authored, and a plan they can read and edit before anything runs. Usable in one sitting, by someone who has never programmed a robot, while being observed.
Used in thesis work · EventMakAR
Turns a scenario written in plain English into assignments drawn from the user's own motion library. Output that does not hold up structurally is refused and regenerated rather than passed downstream, and the result is always shown to a person before it reaches hardware.
Used in thesis work · nursing video
Spatial overlays registered to the room, driven by the same event file a non programmer authored in the browser. The runtime parses the file and drives the physical characters, so authored stories run without anyone opening the Unity project.
Used in EventMakAR
The centre of the fleet: takes an approved plan, splits it per robot, dispatches commands and watches what comes back. Also where an emergency stop lives, which matters when the robots are on the floor of a room with participants in it.
Used in thesis work · the motion study
Knowing where each robot actually is, not where it was told to be. Without accurate position you cannot tell whether a robot performed the motion a person authored, which makes the whole study unanswerable. My first localization approach was not accurate enough and I replaced it.
Used in thesis work · the motion study
The robots themselves. Microcontroller firmware driving motors and servos over wifi, on a chassis built for roughly thirty dollars a unit so a study can use a fleet rather than one robot.
Used in the motion study · thesis work
What a session leaves behind: authoring logs, execution traces, questionnaire responses and interview recordings, captured so that what someone said while authoring can be read against what they reported once they were done.
Used in thesis work · the motion study
Turning sessions into findings: counterbalanced within subjects comparison, descriptive and inferential statistics, and qualitative coding of what participants said, plus the figures that report it.
Used in thesis work · the motion study
A nine module forensic pipeline in Python, evaluated on ProGAN, BigGAN and CycleGAN datasets. Feature selection at k = 20 was optimal for AdaBoost, reaching 87.20% mean AUC. Delivered with a written forensic report.
Explainable machine learning for web content filtering, with a human in the loop feedback mechanism so that filtering decisions can be inspected and corrected rather than just accepted.
Bachelor’s final year project: a working three axis CNC machine built from scratch: mechanical assembly, motor control and G code path planning. My first experience of debugging a system where the bug can be mechanical.
London Metropolitan University · 2022–2023
A deep learning intrusion detection system for classifying network attacks, built as a graduate course project.
BizGees Ltd. · United Kingdom (remote)
Accessibility testing, and working with developers to resolve the usability issues it surfaced.
Danfe Solution Pvt. Ltd. · Kathmandu, Nepal
Resolved technical and security issues; supported firewall, VPN and vulnerability reporting.
Graduate and undergraduate course projects, kept for completeness rather than as research output.