About and Bio

Four years inside a medical device plant, asking whether XR survives contact with it.

Most industrial XR research evaluates the experience. I spent the doctorate on the part that usually gets skipped: whether the mechanical engineers who are meant to use it can build the thing themselves, without me in the room, and keep doing it after the project closes.

Portrait of Sahir Sharma.

Sahir Sharma · TUS Midlands Midwest

Doctorate

Technological University of the Shannon, Midlands Midwest, Athlone

Years 3 to 4

Co-funded by Boston Scientific Galway and the Research Ireland ADAPT Centre

Before

MSc Computer Science, Trinity College Dublin · University Teacher, University of Limerick

Professional Narrative

What an industry co-funded PhD demands that a standard studentship does not

The project opened in October 2022 with the wrong problem. The partner wanted ergonomic testing of manual workstations simulated in XR. Four months of device benchmarking and on-site observation showed that ergonomic testing happened too rarely to justify the investment, while a different activity recurred constantly: equipment design reviews, where cross-functional teams argue about a machine that does not exist yet, using screenshares, 3D printed parts and cardboard floor layouts.

Re-scoping cost time I did not have, and it is the most useful thing I did. It is also the habit that an industry setting enforces. A research question here has to survive contact with people who will simply stop attending if the work does not help them. The decisions that followed came from that constraint: choose a cross-platform interaction toolkit over a single vendor SDK, because the partner will outlive any one headset generation; version the instruction set alongside the software, because the instructions are what actually gets handed over; design the study so the data still holds if half the participants cancel.

Commercial viability as a research constraint

A lot of XR work in manufacturing reports a demonstration and calls it adoption. The partner's engineering lead put the turning point plainly: the gimmicky use of XR was over, and adoption needed to be mainstreamed. That reframing shaped the last two years. It meant testing a polished commercial platform beside my own tooling and reporting where the commercial product was better. It meant costing licence models and infrastructure instead of assuming somebody else would. It meant designing a training pathway and a power user tier, because a capability that depends on one researcher is not a capability.

It also meant being deliberate about ownership. Working on live equipment designs inside a regulated manufacturer, the research had to be publishable without exposing partner intellectual property: anonymised participants, redacted geometry, generalised use cases, methods and instruments released openly while the designs stayed inside the plant. That is a skill in itself, and it is one reason the work could be published at all.

Version churn, and why four templates took twelve builds

The least glamorous finding of the whole doctorate is that long term support is hard to promise in this space. Over four years, every Unity release, XR Interaction Toolkit update, Meta SDK revision and even Meta Link update broke interoperability somewhere in the chain. More than twelve template versions were produced as a result. Staying close to the state of the art and staying stable pull in opposite directions, and a manufacturer needs to hear that before it plans a rollout, not after.

The transfer viva that changed the thesis

Midway through, my examiners said the work read as an industry report rather than a scientific contribution. The response was not to soften the industrial side but to sharpen the question underneath it. The capstone study stopped asking whether XR helps a design review and started asking whether a non-programming engineer can learn to author one, measured step by step, supervised and then unsupervised, with task load and usability instruments attached to each phase and critical incident analysis across the recordings. That is the difference between a delivery and a finding.

Where this leaves me

I am comfortable in the seam between the two worlds. I can specify a work package with a project manager and defend a methodology with an examiner. I can hand a mechanical engineer a headset and a document and then say nothing while they struggle with step four, because that struggle is the data. I am now looking for postdoctoral positions and industry partnerships where I can keep doing that at a larger scale. Get in touch if that is useful to you.

Core Competencies

The same project, read two ways

Each row is one capability the doctorate demanded. The left column is what it looked like to an examiner. The right is what it looked like to the partner's project board.

Framing
Research questions, aims and objectives derived from a literature gap and defended at transfer
Work package definition, scope negotiation and re-scoping with a project board when the original use case proved marginal
Methodology
Mixed method designs across three studies: reflexive thematic analysis, critical incident technique, Delphi consensus, focus groups
Evaluation plans that fit production schedules, room bookings and participant availability, with contingency for no-shows
Instruments
SIM-TLX, System Usability Scale, BFI-2-S, Likert batteries, structured and semi-structured protocols
Scorecards leadership can read, benchmarked against commercial alternatives and tied to time and cost
Engineering
Reproducible experimental apparatus, with telemetry capture written into the application itself
Four production-grade Unity templates with versioned instruction sets, plus a rebuild cadence forced by vendor SDK churn
Data
Roughly 3 TB multimodal corpus: multi-angle video, 360 capture, timestamped headset and hand telemetry, transcripts, coded extracts
Data handled inside enterprise policy, anonymised at source, stored on institutional infrastructure rather than partner systems
Ethics and IP
Research ethics approval, informed consent, GDPR-compliant anonymisation, supervisory data authority correctly identified
IP-aware publication strategy: methods and instruments released, partner geometry and designs redacted
Dissemination
Peer reviewed publication at ACM IMX and ACM MMSys, an accepted workshop position paper, journal reviewing, conference posters
Site workshops, demonstrations to directors and external bodies, an award-winning partner nomination, and hands-on training for engineering teams
Sustainability
Limitations, generalisability, and a research agenda that follows from them
Power user identification, learning and development pathway, site to site capability transfer and a written future state plan

Research Principles

Five positions the work argues for

These are not preferences. Each one came out of a result, and each is tied to the method that produced it.

Measure the creation, not only the experience

Industrial XR studies overwhelmingly evaluate the finished experience. The capstone study instead instrumented the authoring process itself, capturing task load after every step with SIM-TLX (Harris, Wilson and Vine, 2020) and workflow usability with the System Usability Scale (Brooke, 1996). The comparison between supervised and unsupervised sessions is only possible because the unit of analysis moved.

Supervision can hide the flaws it is compensating for

Watching an expert guide a session makes a workflow look more usable than it is. Comparing 14 supervised sessions against 13 unsupervised ones, with critical incidents coded from 360 degree recordings using Flanagan's technique (1954), separated what participants had genuinely learned from what my presence had been quietly fixing.

Documentation is an instrument, not an afterthought

Instruction sets were revised from observed breakdowns rather than from my assumptions about what was difficult. Where a reader stopped, backtracked or improvised told me more than any retrospective question, and the revised document became the independent variable in the next phase.

Co-create with the people who inherit the capability

Stakeholders shaped the research question, the templates and the instructions, and the ethics committee was right to ask about the power imbalance that creates. Treating adoption as a community of practice problem (Lave and Wenger, 1991) rather than a training problem is why the tooling was still in use after I stopped maintaining it.

Publish the limits, including the commercial ones

Reviewers asked for quantitative validation the first study did not carry, and my examiners said the work read as an industry report. Both were right. The vendor evaluation reached a split recommendation that was not in my interest to write. Reporting where the in-house route loses is what makes the rest of it credible.

References: Brooke, J. (1996) SUS: A quick and dirty usability scale. Flanagan, J. C. (1954) The critical incident technique, Psychological Bulletin 51(4). Harris, D., Wilson, M. and Vine, S. (2020) Development and validation of a simulation task load index (SIM-TLX). Lave, J. and Wenger, E. (1991) Situated Learning. Soto, C. J. and John, O. P. (2017) Short and extra-short forms of the Big Five Inventory-2.