Built Different. Literally.
I build AI systems that work in the real world — not just in research papers. Currently finishing my Master's in AI at Sheffield, and thinking seriously about what to build next.
I'm Dimasha Fernando — a Sri Lankan engineer studying AI at the University of Sheffield, and someone who has never been satisfied with just building things that work. I want to build things that matter.
My placement at Airedale by Modine taught me something most CS courses don't: real AI is messy, constrained, and deeply physical. I designed an end-to-end Edge AI system for detecting refrigerant leaks in industrial chillers — from raw sensor data through to a deployed product. That experience changed how I think about building.
Outside the lab, I'm drawn to the collision of technology and entrepreneurship. I've led a £32,500 committee budget as Chair of University of Sheffield's Societies Committee, run cultural events as President of the Sri Lankan Society, and attended more startup networking events than I can count — because I believe the best products come from engineers who also think like founders.
Basketball player. Curious by default. Always thinking about the next problem worth solving.
We are at an inflection point. AI is no longer a research discipline — it's an engineering material. Like electricity was to the 20th century, AI will quietly rewire every industry in the 21st. Most people are still treating it like a feature.
I believe the visionaries who matter will be those who can hold two things simultaneously: the technical depth to understand what's actually possible, and the founder's instinct to ask — but should we, and for whom?
Built end-to-end: from sensor data collection and research through to anomaly detection model design, validation, and deployment on-device. This work became part of the CoolingAI™ product line and laid the groundwork for future AI development within the company.
LLMs hallucinate rather than admit uncertainty. This paper investigates whether post-training alignment methods can teach models to abstain from questions they can't answer reliably — evaluated against the AbstentionBench benchmark across four small-to-mid-size models. My contribution focuses on Prompt Tuning: automatically optimising discrete text prompts to steer abstention behaviour, using an iterative critique-rewrite loop (CriSPO) served via vLLM on HPC. A recurring finding across approaches is that naive interventions tend to act as blunt abstention dials rather than input-conditional discriminators.
Built a multi-tiered translation algorithm in Python that converts UI test cases between Android frameworks — achieving a 95% translation rate. Involved deep research into statistical phrase-based translation.
We're waiting for a dramatic moment that may have already passed quietly. On AI consciousness, algorithmic manipulation, and why the real threat isn't a robot uprising — it's us forgetting how to think.
We exist not as individuals but as a collective. It is this collectiveness that underpins our existence.
How would you know if your friend or family is them or a hyper-realistic android with their memories.
I'm open to full-time roles in AI/ML engineering, particularly where there's a product-focused culture and hard engineering problems. Download my CV or reach out directly.
Download CV →If you're working on something ambitious at the intersection of AI and real-world products and you're looking for a technical co-founder or collaborator — I'd genuinely love to hear about it.
Start a conversation →I'm actively exploring startup ideas — particularly in AI, Behavioural Analysis and anything tech. If you run a programme or back early-stage technical founders, let's talk.
Reach out →Or just say hello
dimasha@dimashaf.com