Master of Computer Science student at the University of Melbourne, passionate about AI, machine learning and scalable software engineering.
Open to internships and graduate roles
Yash Dabas
The University of Melbourne
Based in
Melbourne, AU
Studying
MSc Computer Science
Focus
AI / ML Engineering
Status
Open to opportunities
I started in computer science because I wanted to build things that worked. I moved toward artificial intelligence because I wanted to build things that could work out what to do. Those are different problems, and the second one is far more interesting.
My undergraduate degree was computer science engineering with a specialisation in AI and machine learning. I am now completing a Master of Computer Science at the University of Melbourne, where the coursework has run from advanced database systems and web security through to the ethics of putting automated decisions in front of real people. This semester is spatial data management, AI planning for autonomy, social computing and mobile systems programming.
The through-line in my work is a suspicion of results that have not been stress-tested. A model that scores well on a clean held-out split has told you very little. What matters is what it does under compression, under distribution shift, under inputs nobody curated. I build the evaluation harness before I get attached to the number.
The goal
To work on AI systems at the point where research stops being a notebook and starts being infrastructure other people depend on.
The approach
Understand the problem before reaching for the architecture. Measure honestly. Ship something that survives contact with inputs it was not designed around.
Five projects spread deliberately across research, full-stack, systems and embedded, chosen to cover different failure modes rather than repeat one comfortable pattern.
A platform for running technical interviews online without the candidate being able to cheat, and without treating them like a suspect. Signal-based integrity monitoring paired with timed coding games that actually test how someone thinks.
A role-based platform connecting graduates, students and faculty: directory, events, opportunities and mentorship, built on a normalised relational schema with security treated as a requirement rather than a patch.
A clinical operations system covering patient records, appointment scheduling, staff rostering and billing, designed around a domain model where the hard constraints are enforced in one place.
An embedded system that tracks bay occupancy in real time and automates barrier access, running a deterministic state machine on constrained hardware where the sensors are unreliable by nature.
A records system written in C++ without a framework or database underneath it, as an exercise in owning the class hierarchy, the persistence format and the memory, deliberately.
Tools are not achievements. These are the ones I reach for, grouped by the kind of problem they actually solve.
Areas of active work rather than a publication list. Where I read, reproduce and argue with the literature.
How global self-attention changes what a model can see. A convolution reads a neighbourhood; a transformer relates every patch to every other patch, which makes it structurally suited to noticing that two distant regions of an image disagree with each other. The open question is what that costs in data efficiency, and whether transfer learning genuinely closes the gap or just hides it.
Detection, classification, and the gap between benchmark performance and behaviour on real inputs. Most of the difficulty in applied vision is not the model. It is compression, lighting, occlusion, distribution shift, and everything else a curated dataset quietly removes before you ever see it.
Training dynamics, regularisation, and the discipline of evaluation. A result that cannot be reproduced under a fixed seed, a fixed preprocessing path and a fixed split is not yet a result. It is an anecdote with a chart attached.
The statistical grounding underneath the architectures: bias-variance behaviour, what a validation split can and cannot tell you, and why the choice of metric silently decides which model appears to win.
Reading, reproducing and questioning published work. The reproduction step is where most of the learning happens, because the details that decide whether a method works are rarely the ones highlighted in the abstract.
Education and the projects that ran alongside it, deliberately spread across research, systems, web and embedded work.
The University of Melbourne
In progressCompleted so far: advanced database systems, web security, and the ethics of deploying AI in systems that affect people. This semester moves toward autonomy and systems that run in the world rather than in a notebook.
Proctored technical interviews and assessment games
Designed and built a platform for running technical interviews remotely without the candidate being able to cheat, using instrumentation of the session rather than surveillance of the person.
Minor specialisation in AI & Machine Learning
Four years of computer science fundamentals with a formal specialisation in artificial intelligence and machine learning, built alongside a steady run of systems and full-stack projects.
Alumni Portal · Hospital Management · Smart Parking · C++ Systems
A run of projects deliberately spread across the stack, covering relational backends, object-oriented domain models, embedded firmware and low-level C++, chosen to hit different failure modes rather than repeat one comfortable pattern.
An agenda, not a portfolio. These are the areas I am actively working in now, where the interesting problems have moved, and where my next projects are going.
Systems that plan, call tools and recover from their own failures. The interesting problem is not capability but reliability: how an agent knows it is stuck, and what it does next.
Production surfaces on top of language models, where latency budgets, streaming, cost per request and graceful degradation matter as much as prompt quality.
Grounding generation in real corpora. Chunking strategy, hybrid retrieval and reranking decide whether a system cites the right passage or a confident wrong one.
Standardised tool interfaces that let models act on real systems, with the boundary between capability and authority drawn deliberately rather than by accident.
Decomposition, delegation and synthesis across specialised agents, plus the orchestration and verification layers that keep independent reasoning from compounding into confident nonsense.
Diffusion and transformer architectures for synthesis, approached from both directions: building generative systems, and detecting what they produce.
Open to internships and graduate roles. Whether it is a role, a research collaboration or a problem you think is unsolvable, I read everything, and I reply.