Explore a range of computer science research internships to complete as part of your degree during the semester break.
The following internships listed are due to take place across the Summer break.
Applications open 15 September and close at midnight on 4 October 2026.
Supervisor: Athman Bouguetteya
Eligibility:
Good Programming Skills (Python); Data Handling and Management Skills; Interest in Hardware Instrumentation and Experimental Measurement; Experience with Mobile App Development or Sensor Data Logging will be a plus.
Project Description:
The Internet of Things connects billions of battery-powered devices with limited battery capacity. Crowdsourced Energy-as-a-Service addresses this constraint through short-range wireless energy exchange between nearby devices. The exchange takes place in bounded venues where people gather, called microcells, such as cafes and libraries. A provider advertises a committed amount of energy, and a consumer typically receives less. Such energy loss determines whether a service commitment is fulfilled. The loss depends on the relative movement between the two devices, the venue layout, and the device state. Current platforms request energy through fixed forms and record transfer conditions coarsely. This project develops an agent-based energy service app for mobile devices. The agent interprets a user request in natural language, discovers nearby providers, and monitors the delivery as the devices move. The app logs the relative movement and the resulting energy loss of every transfer. The student will build the app, collect real measurements, and train models that predict loss from peer-to-peer mobility.
Requirement to be on campus: Yes *dependant on government's health advice
Supervisors: Prof Athman Bouguettaya
Eligibility:
Project Description:
Falls are a major cause of injury, hospitalisation and loss of independence among older adults. As more people age at home, there is a growing need for technologies that can predict increasing fall risk before a fall occurs. Passive infrared (PIR) detectors, door contacts, bed mats and radar can continuously record activity, sleep and movement between rooms without cameras or wearables. However, most in-home systems only detect signal only after a resident has fallen. While clinical fall-risk assessments exist, a critical challenge remains: they are made once or twice a year, but the changes before a fall unfold over weeks. Since these changes are gradual and differ between people, a fixed threshold cannot separate a real decline from a quiet week. AI makes this tractable, because sequence models can learn from unlabelled data what normal looks like per person. To address this gap, this project explores an AI-driven Digital Fall Biomarker, which infers a resident's unobservable risk state with a hidden Markov model, and visualisation of the resulting risk trajectory.
Requirement to be on campus: Yes *dependent on government’s health advice
Supervisor: Dr Hong Jin Kang, Dr Rahul Gopinath
Eligibility Criteria: Good or strong programming ability
Project Description:
Real-world software carry implicit security rules: user permissions have to be checked before sensitive operations, or input must be validated before it's trusted. However. these rules are rarely written down, so the same mistakes are made again as a project evolves.
In this project, we will develop an agentic system that studies a project's past security fixes, learns the rules behind the fixes, and automatically synthesizes a code-scanning query in CodeQL to detect violations of same rule.
The project will provide experience with large language models, AI agents, and experimental evaluation.
Requirement to be on campus: No
Supervisor: Dr Rahul Gopinath, Dr Hong Jin Kang
Eligibility:
Project Description:
Students are the primary user of LLM, as it can provide explanations and feedback quickly. With the growing adoption of tutor and long horizon agent, the increasing reliance raises concerns about misleading or overly direct answers that limit deeper comprehension. Different LLM personas and styles may significantly influence meaningfulness and effectiveness of learning. There is currently no standard benchmark for evaluating LLM personas, most existing benchmarks focus on task accuracy or reasoning ability.
To address this gap, to develop a benchmark dataset which consists of tutoring tasks and an evaluation metrics is possible. Systems can interact with these tasks, and their performance will be evaluated using metrics.
By providing an open-source benchmark that allows researchers to evaluate systems and models, this work also helps to enhance the trustworthiness of LLM and tutor agents for student, reduce negative learning outcomes, and support the design of more effective AI tutors.
Requirement to be on campus: No
Supervisor: A/Prof. Anusha Withana, Dr. Zhanna Sarsenbayewa, Aurélia de Silva
Eligibility:
Essential:
Desirable:
Project Description:
3D models are increasingly created, shared, and modified across tools and design stages. Many people want to edit the 3D models for their own usage, but, preserving the shared model’s editability, internal structure, and intended behavior remains challenging. We plan to explore new representations and systems for portable, editable 3D models.
During this internship, you will develop methods for representing, processing, modifying, evaluating, or maintaining structured 3D models. Working closely with PhD researchers and supervisors, you will formulate a focused research question, assisting with implementing and evaluating a solution, and report the findings. If we manage to get good results, we aim to publish the work in computer graphics or HCI conference.
This line of research has already led to papers accepted at and submitted to A* conferences. The internship extends this work. For resulting publications, authorship and author order will reflect the nature and extent of your contributions.
Requirement to be on campus: Yes *dependent on government’s health advice
Supervisor: Dr. Anusha Withana, Praneeth Perera
Eligibility:
Project Description:
Braille provides an essential way for people with vision impairment (PVI), including those with low vision or blindness, to access written information through touch. However, refreshable Braille displays can be costly, bulky, or difficult to access. Limited research has explored whether compact electro-tactile feedback can present clear and comfortable Braille patterns directly through the skin.
This project aims to explore the feasibility of using electro-tactile (ET) feedback for Braille reading. We will begin with simple tactile patterns and gradually investigate how combinations of patterns could represent Braille characters. The student may contribute to prototype development, signal design, software integration, and early-stage evaluation.
This project develops on papers published at A* conferences. If we manage to get good results, we aim to publish the work in an HCI conference. The authorship and author order will reflect the nature and extent of your contributions.
Requirement to be on campus: Yes, in-person laboratory work is required.*dependent on government’s health advice.
Supervisor: Dr. Anusha Withana
Eligibility:
Project Description:
Large structures are hard to move. Imagine something like a solar panel used in satellites. What if such large objects could be 3D printed small enough to fit in a backpack, then deployed into something meters across?
This project looks at how we design and fabricate shape-changing deployable structures. You will help build a design tool that generates printable geometries such as chambers, creases, and folds, and predicts the shape they take once deployed. You will then print your own prototypes in our lab, deploy them, and measure how closely the real shape matches the design.
You will get hands-on experience in computational geometry, digital fabrication (3D printing), and rapid prototyping, and contribute to ongoing research on deployable structures. No prior experience in this area is needed. Curiosity and a willingness to build things matter most.
This project develops on papers published at A* conferences. If we manage to get good results, we aim to publish the work in an HCI/Graphics conference. The authorship and author order will reflect the nature and extent of your contributions.
Requirement to be on campus: Yes, in person laboratory work is required, *dependent on government’s health advice.
Supervisor: A/Prof Chang Xu
Eligibility:
Project Description:
This project focuses on building scalable data and training infrastructure for large-scale machine learning. The successful candidate will work on systems supporting high-throughput data ingestion, distributed data processing, efficient storage and I/O, multi-node GPU training, and reliable long-running training workloads.
Research topics may include distributed training with PyTorch and NCCL, GPU cluster scheduling, high-performance data pipelines, object storage and local NVMe caching, large-scale dataset formats such as Parquet/Arrow, fault tolerance, checkpointing, and performance profiling across storage, network, CPU, and GPU resources.
The project is particularly suited to candidates with hands-on experience in large-scale ML systems, distributed systems, high-performance computing, or data infrastructure.
Requirement to be on campus: Yes *dependent on government’s health advice.
Supervisor: A/Prof Chang Xu
Eligibility:
Project Description:
The system uses a single RGB-D camera to scan a standing person from the front, side, and back, and computes postural metrics such as shoulder height difference, and head tilt and forward head position. Abstractly, this is a problem of estimating 3D anatomical landmarks from depth and color data, then measuring small asymmetries, on the order of millimeters and degrees. The difficulty is that the quantities of interest are close in magnitude to the noise: sensor error, landmark localization error, postural sway, and repositioning between views all sit in the same range.
Requirement to be on campus: Yes, *dependant on government's health advice.
Supervisor: Dr Clément Canonne
Eligibility: Having taken COMP3027 or COMP3927 (or equivalent) with a DI or HD, solid background in discrete mathematics.
Project Description:
This project will focus on understanding the power, limitations, and applications of quantum pseudodeterministic algorithms, as introduced by Aaronson, Gur, and Li, that is, quantum algorithms which, in spite of the inherent randomness of quantum computation, can be made to (nearly) always output the same answer.
Requirement to be on campus: No
Supervisor: Dr. Rahul Gopinath
Eligibility: Strong programming skills and familiarity with software testing or systems programming. Knowledge of formal grammars, fuzzing, compiler optimization, assembly, or Rust is desirable but not required.
Project Description:
Grammar-based fuzzers automatically generate structured inputs for testing software such as parsers, interpreters, and compilers. F1 achieves high throughput by compiling grammars directly into assembly, while recent approaches such as FANDANGO-RS (High-Performance Generation of Constrained Inputs) combine compiled grammar representations with evolutionary algorithms to generate inputs satisfying semantic constraints. This project will investigate whether supercompilation (partial evaluation) and/or superoptimization can further improve F1’s generation performance and extend its capabilities to selected semantic constraints. The student will develop optimizations for generated assembly, explore techniques for incorporating constraints directly into input generation, and compare the resulting system against FANDANGO-RS. Evaluation will consider generation throughput, input validity, structural diversity, and optimization overhead across representative grammars and constraints. Outcomes include an open-source prototype, reproducible benchmarks, and a research presentation, with potential for publication.
Requirement to be on campus: No
Supervisor: Dr Rahul Gopinath
Eligibility: This project requires a fast learner, who is comfortable with doing literature search, understanding the algorithms, along with the implementation.
Project Description:
Parsers are one of the key interfaces between computers and human beings. A general parser takes a specification (typically provided as a grammar) the input string and generates a derivation tree. A key difficulty when using parsers is that in many languages, specific operators have specific precedence. For example, multiplication (*) typically has higher precedence than addition (+). A typical way to account for this is to try to encode this precedence in the grammar itself. However, this is error-prone and has led to a bad reputation for general context-free parsers. The alternative is to extract the first parse tree and transform the tree to the derivation tree with the correct precedence bracketing. LaLonde et al showed how to do this for simple languages (Wilf R. Lalonde and Jim Des Riviers 1981).
This project will involve (1) doing the background literature review (2) understanding and implementing LaLonde’s algorithm (3) identifying and applying this algorithm on common programming language patterns that typically are solved with handwritten precedence parsers and evaluating the effectiveness.
(In your application, please indicate if you are interested in a Honours application with the supervisor).
Requirement to be on campus: No
Supervisor: Dr Rahul Gopinath
Eligibility:
Strong programming skills and familiarity with Linux command-line tools. Knowledge of automata theory, formal grammars, fuzzing, or software testing is desirable but not required.
Project Description:
Command-line programs such as GNU find, GCC, and Clang accept complex combinations of options, arguments, and flags. Effective fuzzing must explore these combinations, but existing fuzzers often waste effort generating invalid command lines that are immediately rejected. This project will investigate whether grammar inference can automatically recover the structure and constraints governing valid command-line options. The student will implement an inference approach based on the TTT algorithm, incorporating prefix queries and examples extracted from instrumentation and documentation to learn option grammars for widely used command-line programs. These inferred grammars will then guide the generation of valid and diverse command lines for fuzzing. Evaluation will compare the approach against conventional fuzzing techniques using measures such as command validity, option-combination coverage, program coverage, and bug-finding effectiveness.
Requirement to be on campus: No
Supervisor: Prof Jinman Kim
Eligibility: WAM≥75 and Undergraduate candidates must have already completed at least 96 credit points towards their undergraduate degree at the time of application.
Project Description:
Hospital in the Home (HITH) is an emerging model of healthcare delivery that provides acute hospital-level care in patients' homes as an alternative to inpatient admission. Growing pressures on healthcare systems, including population ageing, increasing chronic disease burden, and hospital capacity constraints, have accelerated interest in HITH services.
Advances in digital health technologies, including wearable sensors, remote patient monitoring, virtual care, and artificial intelligence (AI), create opportunities to develop new models of care beyond traditional hospital delivery. These technologies can support more personalised, proactive, and connected healthcare through continuous monitoring and intelligent clinical decision support.
In collaboration with the Nepean Hospital, this project will develop digitally enabled HITH models for selected clinical services across a large and geographically diverse region where HITH is essential. Potential services include virtual wards, AI-assisted decision support, and home-based rehabilitation.
Students with backgrounds in digital health, virtual care, or AI are strongly encouraged to apply.
Requirement to be on campus: Yes *dependent on government's health advice.
Supervisors: Dr Sichao Li, Prof Mary Lou, Dr Zhengyi Yang
Eligibility:
Project Description:
Large language models are increasingly used as AI tutors, but a tutor can appear helpful while unintentionally doing too much of the reasoning for the student. This project will investigate how such reasoning disclosure can be systematically identified and evaluated.
The student will contribute to an ongoing research project developing a benchmark for auditing AI tutoring dialogues. They will help construct reasoning-based evaluation tasks from existing educational datasets, represent key intermediate reasoning steps, and analyse whether different AI tutors reveal these steps directly, indirectly, or appropriately leave them for the learner. The project will also explore using large language models as automated auditors and compare their judgments with human annotations.
The internship will provide hands-on experience in LLMs, NLP, AI in education, benchmark design and empirical evaluation, with the opportunity to contribute to a research dataset, experimental results, and potentially a research publication.
Requirement to be on campus: No
Supervisor: Dr. Sichao Li, Prof. Kimberlee
Eligibility: Suitable for students in Computer Science, Software Engineering, Data Science, or a related field. Strong Python programming is required. Familiarity with machine learning, large language or vision-language models, data processing, or human-subject annotation is desirable.
Project Description:
Current vision-language models can describe what may happen next, but less is known about whether they understand what a reasonable person should do over different time scales. This project extends NoRA, a benchmark for grounded normative reasoning in first-person video, with a temporal dimension. For each NoRA clip, the student will construct a canonical set of candidate actions and obtain human judgments of whether each action is reasonable immediately, in the short term, and over longer horizons. The student will build an annotation interface that presents the video, fixed scene facts, candidate actions and temporal bands while enforcing blinded, independent judgments and recording agreement. The project will develop data-processing and quality-control pipelines, run a pilot annotation study, analyse temporal shifts in reasonableness, and prepare a reproducible Temporal NoRA dataset extension. Stretch goals include multi-select evaluation questions and baseline testing of multimodal language models.
Requirement to be on campus: No
Supervisor: Dr. Sichao Li, Dr. Zhengyi Yang, Prof. Jiangshan Yu
Eligibility:
Essential:
Desirable:
Project Description:
Geospatial scientific data are often stored as polygons, such as bushfire extents, but current AI systems have limited ability to retrieve and reason over these shapes using natural-language queries. This project contributes to PolyGeoAlign, an ongoing study of language-aligned representations for spatiotemporal polygons. The student will help build a benchmark using Australian bushfire extent data, with a particular focus on human annotation of subjective spatial concepts such as irregular shape, narrow corridors, similar morphology, and temporal growth patterns. The student will also assist with query design, annotation-quality analysis, inter-annotator agreement, data preparation, and lightweight retrieval baselines. The project offers practical research experience in GeoAI, multimodal/representation learning, spatial data, information retrieval, and reproducible benchmark design. Strong outcomes may contribute to an open research benchmark and research publication.
Requirement to be on campus: No
Supervisors: Dr Sri AravindaKrishnan Thyagarajan, Dr Clement Canonne, Prof Qiang Tang
Eligibility: Third year or later undergraduate in Computer Science, Mathematics or Physics. Comfortable with linear algebra and with reading and writing mathematical proofs. No programming and no laboratory access are required. WAM above 85.
Project Description:
Imagine handing someone a program they can run once, on an input of their choosing, after which it is useless to them. On an ordinary computer this is impossible unless you also hand over tamper proof hardware. Quantum information cannot be copied, so a quantum state might enforce single use on its own. The research so far pulls in two directions. There are strong results showing it cannot be done for general programs, and there are constructions that do work, in restricted settings or by assuming something extra. Nobody has put the two sides side by side. This project asks where the line between them actually falls, by working through the arguments for impossibility, the constructions that get around them, and the choices of definition that decide which side a result lands on. The output is a report placing the known results in one common framework.
Requirement to be on campus: No
Supervisors: Sri AravindaKrishnan Thyagarajan, Liyi Zhou and Qiang Tang
Eligibility: Comfortable with linear algebra and probability, and willing to read careful mathematical arguments. Some Python is useful for running the estimation tools, but no specialist hardware and no machine learning are needed. WAM greater than 85.
Project Description:
Every system built to survive future quantum computers rests on a mathematical problem that experts believe is too hard to solve. The well known problems have been studied for decades. Newer systems, the kind that let a group of people sign something together, or let you prove a statement without revealing why it is true, rest on newer problems invented for the purpose, and these appear faster than anyone checks them. A public catalogue of them opened in 2026 and is admittedly incomplete. One entry has already been broken and others turned out to be restatements of each other. The catalogue records how the problems relate. It says nothing about how large the keys must be. This project works that out, extending the catalogue and using open tools to estimate how hard each problem really is, then comparing that against the key sizes real systems have chosen.
Requirement to be on campus: No
Supervisors: Wei Bao
Eligibility: Knowledge in federated learning and unlearning
Project Description:
In large units, instructors often manage hundreds of students and multiple assessments, creating significant challenges in providing timely and meaningful feedback. Tutors may require considerable time to review assignments, and delays in feedback reduce its effectiveness for learning improvement. Students also frequently request clarification on grading decisions, further increasing tutor workload.
Federated semi-supervised learning enables multiple clients to collaboratively train machine learning models using limited labeled data and abundant unlabeled data, while keeping local data private. However, when a client or its data needs to be removed, existing federated unlearning methods are mainly designed for fully supervised settings and may not effectively remove the influence propagated through unlabeled data training.
This project will investigate federated semi-supervised unlearning, with the aim of developing an initial framework for efficiently removing the influence of selected clients or data without retraining the entire model from scratch. The project will first establish a federated semi-supervised learning environment, examine how forgotten information may affect both labeled and unlabeled data, and explore preliminary unlearning strategies. The effectiveness of the proposed approach will be evaluated against standard retraining and existing federated unlearning baselines.
Requirement to be on campus: Yes *dependent on government’s health advice.
Supervisor: Dr Liyi Zhou
Eligibility: Essential: strong programming ability and mathematical maturity. Suitable backgrounds include computer science, software engineering, mathematics or related disciplines. Desirable: discrete mathematics, algorithms, security or cryptography. Python and careful technical reading are important; formal-methods experience is helpful but not required.
Project Description:
Cryptographic papers can fail because a security claim, assumption or proof step is incomplete. This project will explore whether an AI research agent can discover such flaws from first principles. The student will build a curated benchmark of public cryptographic claims with known flaws or counterexamples, then develop an agent that extracts proof obligations, challenges assumptions, searches for edge cases and produces checkable evidence. The study will separate plausible criticism from validated flaws through structured rubrics, counterexample construction and, where feasible, lightweight formal or executable checks. It will evaluate unseen papers and flaw types, document false positives and identify transferable reasoning strategies. Deliverables are a prototype, benchmark and rigorous evaluation report, with potential for an open-source artifact or publication. The work uses public research materials only and does not target deployed systems.
Requirement to be on campus: No
Supervisors: Dr Liyi Zhou
Eligibility: Essential: confidence with Python and Git and completion of at least one programming, software engineering, cybersecurity or AI unit. Desirable: Linux/containers, LLM agents, program analysis or vulnerability research. Careful experimentation and validation matter more than prior security expertise.
Project Description:
AI coding agents can inspect software, but they often repeat mistakes and produce plausible reports that do not reproduce. This project will build a minimal, from-scratch EvoHunt-style prototype that learns an external auditing playbook from grounded successes and failures. The student will assemble a small benchmark of public, source-available vulnerabilities, run agents in isolated containers, require executable evidence, and compare an empty procedure, a fixed expert procedure and an evolved procedure. The research will examine which experiences should be retained, how revisions transfer to unseen repositories, and how to prevent improvement on one case from causing regressions elsewhere. Deliverables are a reproducible prototype, benchmark and evaluation report, with potential to contribute to an open-source release or research paper. No testing of live third-party or production systems will be conducted.
Requirement to be on campus: No
Supervisors: Dr. Mohammad Polash and Muhammad Sajjad Akbar
Eligibility:
Project Description:
With the rapid rise of generative AI, students increasingly rely on AI-generated code to solve programming tasks. While this accelerates problem-solving, it often undermines their confidence in writing code independently. This project addresses this challenge by developing an educational tool that leverages GenAI to strengthen students’ programming skills and computational thinking.
The tool generates intentionally buggy code based on a given problem specification and prompts students to critically analyze it. Learners will be tasked with identifying logical, syntactical, or structural errors, encouraging a deeper understanding of programming concepts. Additionally, the tool guides students to translate problem requirements into step-by-step computational logic, reinforcing their ability to design solutions before coding.
By shifting the focus from passive code consumption to active problem-solving and debugging, the project aims to build students’ confidence, autonomy, and resilience in programming, equipping them with skills crucial for both academic and professional success.
Requirement to be on campus: No
Supervisor: Dr. Sri AravindaKrishnan Thyagarajan
Eligibility: Confident programming in Python, and comfortable reading a technical paper closely enough to pull a formula out of it. Comfortable with logarithms and asymptotic notation. No prior cryptography and no specialist hardware are needed, since the first two weeks are guided reading. WAM greater than 80
Project Description:
Banks, blockchains and government systems do not let one person hold a signing key. They split it, so that several people must agree before anything gets signed. Every method in use today breaks against a quantum computer. Replacements exist, and they behave strangely. One recent scheme handles a thousand signers, while splitting the key of the standard NIST actually adopted currently manages six. Signatures grow from 64 bytes to 13 kilobytes. The papers report different quantities under different settings. This project characterises the techniques behind those numbers. You take each recent construction apart into the handful of choices its designers made, how a share is hidden, how shares are combined, how a failed attempt is handled, and works out which choice drives which cost. The output is a map of the design space, and an explanation of why some schemes scale and others do not.
Supervisor: Dr. Sasha Rubin
Eligibility: All of the work will have some mathematical component, and may include a coding component depending on the topic. This project will suit a student who achieved an HD in an advanced course on Theoretical Computer Science, such as COMP2922.
Project Description:
Planning is part of the symbolic/logic approach to AI which involves finding a finite- state program that tells an agent what to do in every state. The states and possible actions are described declaratively. Please see this link as an example.
One approach to building a planner is to reduce the given planning problem to the satisfiability problem, and implement this as a reduction to SAT solvers or in declarative programming languages such as ASP.
There are a few possible topics for this project, depending on the interest and skill of the student.
E.g., Build a declarative planner for modern (decision-theoretic) solutions such as non-dominated solutions.
E.g., Build a declarative planner that solves “lifted planning” for standard solutions (such as strong or strong-cyclic) by reducing to automated theorem proving.
Requirement to be on campus: No
Supervisor: Dr. Anastasia Isychev
Eligibility:
Project Description:
Neural networks (NN) are widely used in classifying and decision making when the exact logic of the decision isintractable. NNs are therefore often treated as a blackbox component. However, due to this opaque and approximate nature NNs are prone to bugs and do not all always deliver an expected classification or decision.
Testing can discover bugs with respect to such functional properties of NNs but cannot explain the reasons behind them. This exploratory project aims to analyse the inner working of NNs and try to detect patterns that lead to faulty decisions.
Requirement to be on campus: No
Supervisor: Dr. Anastasia Isychev
Eligibility: Background in static program analysis, ability to reason about abstractions, good coding skills
Project Description:
When testing program analyzers using metamorphic testing, a fuzzer runs an analyzer on a pair of input programs and compares the output analysis results, this relation between theanalysis results pairs determines whether a test passes or fails. One frequently used relation on outputs for testing analyzers isequivalence, for instance, the analysis results of programs X andY are expected to be the same if a fuzzer generated program Y by adding dead code to program X. Here, adding dead code is an equivalence transformation, e.g. a way to create a program that is semantically equivalent and thus also equivalent with respectto expected analysis results. Recent work shows that equivalence criteria can be generalised from semantic equivalence. This project aims to explore whetherrelations for such generalisations an d the corresponding equivalence transformations modulo different criteria can bediscovered automatically. The project involves literature research and prototype implementation.
Requirement to be on campus: No
Supervisors: Dr. Anastasia Isychev
MaxEligibility:
Project Description:
Numerical programs present a unique challenge for op4miza4ons: results should be computed quickly but remain accurate enough. Both accuracy and performance depend on many factors, including finite-precision data types (float, double, etc.), concrete arithme4c opera4ons in the expression, and the magnitude of values involved in the computa4ons. However, the most accurate programs are usually slow and the fastest are inaccurate, and the op4miza4on must navigate the trade-off of this mul4ple-objec4ve problem.
Existing work shows that performance can be improved by assigning different finite precisions to variables in one expression while monitoring the effect on the overall accuracy and controlling its loss. The most effec4ve assignments also take into account variables values. Such process of finding suitable values to determine precision assignment is called regime inference.
This project aims to find novel methods for inferring regimes (finding the suitable cutoff values for precision assignment) for a dynamic performance op4miza4on called mixed-precision tuning. The workload will involve reading scien4fic papers to understand exis4ng methods, designing an algorithm for regime inference and implemen4ng it in a tool prototype, and evaluating the findings.
Requirement to be on campus: No
Supervisors: Prof Alan Fekete, A/Prof Uwe Roehm [collaboration with Prof David James in CPC]
Eligibility: Essential: Knowledge of SQL; experience installing software, administration and tuning, and dealing with dependencies etc. Desirable: experience measuring performance of systems; some knowledge of bio-datasets
Project Description:
A research project at the Charles Perkins Centre is developing Protehome, a PostgreSQL-based data management system for integrating biological datasets and experiments. In parallel, our database research group has explored alternative data models and platforms, including graph and document databases. Previous Honours and VRI projects developed an initial set of benchmark queries for comparing these approaches.
This project will extend that work by developing an easy-to-use benchmark suite with broader coverage of different query types. The resulting benchmark will be used to evaluate different database platforms for managing and querying omics data and may contribute to ongoing interdisciplinary research.
Requirement to be on campus: No
Supervisors: Prof Alan Fekete, A/Prof Uwe Roehm (external collaboration with Dr Jim Webber at Neo4j, Dr Michael Cahill)
Eligibility:
Project Description:
Distributed databases must continue to operate correctly even when servers or network connections fail. A commit protocol defines how database nodes coordinate to decide whether a distributed transaction can complete successfully. Designing such protocols is difficult, and subtle bugs may only appear under rare failure scenarios. Formal models provide a precise specification of protocol behaviour and can be automatically checked using verification tools.
Researchers at Neo4j, a leading graph database company, recently proposed a novel commit protocol. An Honours project developed and verified a substantial part of the protocol using the TLA+ specification language. Since then, the protocol has evolved. This project will extend the existing model to cover new protocol features and verify their correctness using a formal model checker.
Requirement to be on campus: No
Supervisors: Prof Alan Fekete, A/Prof Uwe Roehm [collaboration with Prof David James in CPC]
Eligibility:
Essential: Knowledge of SQL; experience installing software, administration and tuning, and dealing with dependencies etc.
Desirable: experience with document databases and/or graph databases; experience measuring performance of systems; some knowledge of bio-datasets
Project Description:
A research project at the Charles Perkins Centre is developing Protehome, a PostgreSQL-based data management system for integrating diverse biological datasets and experiments, such as genome and proteome sequences. In parallel, our database research group has explored alternative approaches using graph and document databases, with prototype systems developed through previous Honours projects.
This project will compare the performance and capabilities of the different approaches using representative data integration and query workloads. The study will assess the strengths and limitations of relational, graph, and document data models for biological data management and may contribute to ongoing interdisciplinary research.
Requirement to be on campus: No
Supervisors: Zhiyong Wang , Yanli Li
Eligibility: WAM > 80. Strong programming skills in implementing foundation models and good knowledge of protein design.
Project Description:
Following the groundbreaking success of large language models for textual data, various protein foundation models have been developed for generative biology, such as protein design. However, it remains an open question how to maximize the value of these foundation models for functional driven protein design. This project aims to leverage the capability of foundation models and protein structure with novel deep learning algorithms for function-guided protein design.
Requirement to be on campus: Yes *dependent on government’s health advice.
Supervisor: Prof Zhiyong Wang
Eligibility:
- WAM >= 80
- Strong computer science background
- Good knowledge of bioinformatics
Project Description:
Recent breakthroughs in Artificial Intelligence (AI) have led phenomenal success in protein structure prediction, which presents immense opportunities to accelerate the innovations in structure based drug discovery. This project aims to develop novel generative AI methods by exploiting the structures of proteins for effective drug design.
Requirement to be on campus: No
Supervisor: Dr Yunke Wang
Eligibility:
Desirable skills:
Project Description:
This project aims to design and develop a wearable teleoperation device for intuitive whole-body control of dual-arm robotic platforms. The system will capture the operator’s upper-body and arm motions and translate them into real-time control commands for a robotic system. The student will contribute to both the mechanical and embedded system development of the device, including mechanical structure design, CAD modelling, sensor integration, electronics prototyping, microcontroller programming, and communication with the robot control system. The project will involve iterative hardware prototyping, system integration, and experimental evaluation of teleoperation accuracy, latency, ergonomics, and usability.
Requirement to be on campus: Yes *dependent on government’s health advice.
Supervisor: Dr Suranga Seneviratne
Eligibility:
Project Description:
Behavioural biometrics are crucial for continuous authentication, particularly as zero-trust architectures gain prominence. Systems can verify user identity in real-time by analysing multimodal data streams – such as keystroke dynamics, mouse movements, gait, and touch patterns. This continuous verification enhances security by detecting anomalies and preventing unauthorised access. However, real-world data is often noisy, incomplete, and subject to environmental variability, posing significant challenges for reliable classification and detection.
To this end, this project offers two potential directions. The first focuses on developing and evaluating JEPA-based models by Meta to learn robust representations from behavioural biometrics data, addressing inconsistencies and improving predictive accuracy.
The second is to explore whether the in-context learning abilities of sensor foundation models can be used for reliable continuous user authentication. This direction is based on the recent work from Google.
Requirement to be on campus: No
Supervisor: Prof. Seokhee Hong, Dr. Amyra Meidiana
Eligibility: Skills Required: Data Structure and Algorithms and Programming (Java, C++, Python, Javascript)
Project Description:
Technological advances have increased data volumes in the last few years, and now we are experiencing a “data deluge” in which data is produced much faster than it can be understood by humans. These big complex data sets have grown in importance due to factors such as international terrorism, the success of genomics, increasingly complex software systems, and widespread fraud on stock markets.
Visualisation is a powerful tool to compute good geometric representation of abstract data to support analysts to find insights and patterns in big complex data sets.
This project aims to design, implement and evaluate new visualisation algorithms for scalable and faithful visualisation of big complex data, to enable humans to find ground truth structure in big complex data sets, such as social networks and biological networks.
These new visualisation methods are in high demand by industry for the next generation visual analytic tools.
Requirement to be on campus: Yes *dependent on government’s health advice.
Supervisor: Imdad Ullah
Eligibility: WAM>75 and Undergraduate candidates must have already completed at least 96 credit points towards their undergraduate degree at the time of application.
Project Description:
This research could investigate an AI-based learning support system that dynamically adapts educational content to students with learning difficulties such as dyslexia, attention-related challenges, language barriers, or low comprehension. The system could personalise the presentation of learning materials by simplifying text, providing audio explanations, using visual summaries, adjusting pacing, generating scaffolded hints, and checking understanding through short adaptive questions. The study would evaluate whether such personalised support improves student comprehension, engagement, confidence, and learning outcomes compared with standard learning materials.
Requirement to be on campus: No
Supervisor: A/Prof Uwe Roehm and Prof Alan Fekete [collaboration with Prof David James in CPC]
Eligibility: Knowledge of SQL; some experience with LLMS; some knowledge of bio-datasets
Project Description:
Protehome is a research database being developed at the Charles Perkins Centre to integrate information from biological datasets and experiments using PostgreSQL. A key challenge is capturing experimental metadata from published research papers so that results can be stored, searched, and analysed alongside existing datasets.
Our research group has previously developed an approach that uses Large Language Models (LLMs) to extract experimental metadata directly from research article PDFs and import the information into Protehome. This project will extend that work by investigating how well LLMs can extract metadata that is presented in figures, diagrams, and tables rather than in plain text.
The student will evaluate different extraction approaches and measure their accuracy on a collection of published research papers. The project provides hands-on experience with LLMs, information extraction, databases, and data management, and may contribute to ongoing research publications.
Requirement to be on campus: No
Supervisors: Dr Clement Canonne and Dr Sri Aravinda Krishnan Thyagarajan
Eligibility: Some experience with measuring database performance; also experience with InnoDB, DuckDB or MongoDB is desirable
Project Description:
JSON is widely used for data exchange and is commonly stored in databases for persistent storage. While the structure of a JSON document often remains stable, some fields, particularly arrays, can grow substantially over time. Examples include IoT sensor readings, activity logs, or event records where new entries are continuously appended.
This project investigates how growing JSON documents affect database performance. In particular, the project will evaluate the impact of increasing JSON document sizes on update and query performance, comparing relational databases with JSON support (eg. PostgreSQL, MySQL/InnoDB, or DuckDB) against document stores such as MongoDB. It will also evaluate whether different indexing strategies can mitigate performance degradation.
The project will use YCSB+IVS, an extension of the Yahoo! Cloud Serving Benchmark (YCSB) recently developed by our research group, to benchmark workloads with increasing value sizes over time. Results from the project will contribute to ongoing research on JSON database benchmarking.
Requirement to be on campus: No
Supervisors: A/Prof Uwe Roehm (external collaboration with Dr Thomas Rueckstiess)
Eligibility: Some experience with measuring database performance; also experience with MongoDB is desirable.
Project Description:
Query optimisation is a critical component of modern database systems. MongoDB traditionally selects query plans using a “First-Past-the-Post" (FPTP) approach, where multiple candidate plans are briefly evaluated before one is chosen. More recently, MongoDB added a new cost-based query optimiser that estimates query costs to guide plan selection.
In previous work, our research group developed an AI-based MongoDB Query Advisor that learns from past query executions using deep reinforcement learning and recommends query hints to improve performance. The advisor relies on its own internal estimates of query selectivity and execution costs.
This project will compare the estimates produced by the Query Advisor with those generated by MongoDB's new cost-based optimiser and investigate how well they predict actual query performance. Results may contribute to ongoing research on AI-assisted database optimisation and query processing.
Requirement to be on campus: No
Supervisors: A/Prof Uwe Roehm(external collaboration with Dr Thomas Rueckstiess)
Eligibility: Some experience with measuring DBMS performance; also experience with LLMs and MongoDB/Couchbase is desirable.
Project Description:
Large Language Models (LLMs) can be used to automatically generate large numbers of database queries for database system testing and benchmarking. Recent work in our research group has extended the SQL Storm approach (VLDB 2025) to support JSON queries for JSON stores such as MongoDB and Couchbase. This project will build on that work and investigate how the approach can be further adapted to relational databases with SQL/JSON support, such as PostgreSQL. An important project aspect is to make sure that the generated queries are semantically consistent between different systems.
We will extend and evaluate the existing LLM-based query generation for a new database and assess its effectiveness for automated testing and benchmarking. The project provides hands-on experience with databases, LLMs, benchmarking, and experimental systems research, and may contribute to ongoing research publications.
Requirement to be on campus: No
Supervisors: Mary Lou Maher, John Vulic, Jacob Elmasry
Eligibility: High-Achieving Students Interested in Engineering and Computing Education
Project Description:
The Faculty of Engineering is planning to implement content specific learning AI Agents in every unit of study beginning in Semester 1 of 2027. While some AI agents have already been deployed in a select number of units, there is not yet a unified understanding of the impact on students’ learning, use, and perspectives on these existing AI-Agents. We are interested in analysing data about the learning approaches and student engagement use in 2026 to inform implementing this approach at scale.
This project seeks to address the question of current and potential impact by collecting information about the implementation of AI agents and their use by students in the Faculty of Engineering, identifying common characteristics of the agents and their transferability to other units, and analysing students’ performance, use and perspectives when they engaged with these agents.
Requirement to be on campus: Yes *dependent on government’s health advice
Supervisor: Dr. Rahul Gopinath, Dr. Simon Poon
Eligibility: Strong programming skills, preferably in Python, and willingness to read clinical decision-logic documents and work with clinical collaborators. Experience with large language model APIs, information extraction, rule engines, or software testing is desirable but not required. No medical background is required.
Project Description:
Breast cancer treatment decisions are made by multidisciplinary teams following detailed, stage-specific decision logic covering initial diagnosis, surgery, and postoperative care. Large language models explain such decisions fluently but cannot be trusted to make them: they blend distinct clinical pathways, overlook missing findings, and retreat into vague generalities. This project will build and evaluate a hybrid decision-aid chatbot in which a deterministic rule engine encoding clinician-authored decision logic selects the recommendation, while the language model is confined to extracting findings from reports and explaining what the rules decided. The student build clinical decision logic as auditable rules, specify the case-routing and completeness-checking layers, and implement a constrained digital helper. Evaluation will use clinician-reviewed case vignettes to measure pathway-selection accuracy, sensitivity to missing or confounding information, answer specificity, and whether the reasons the chatbot gives match the rules that actually fired.
Requirement to be on campus: No
Supervisors: Sri AravindaKrishnan Thyagarajan, Pieter Roffelsen, Tomas Lasic Latimer
Eligibility: A WAM above 85 is required. Knowledge of modular arithmetic and statistics.
Project Description:
The dawn of the quantum era is near, yet much of our cybersecurity still rests on assumptions based on binary computing. In this interdisciplinary project, we search for new quantum safe algorithms by bringing ideas from the modern theory of integrable systems into the discrete world. To this end, we will stress test structures from integrable systems against both state-of-the-art classical attacks and the emerging threats posed by quantum computers. This will be an interdisciplinary project involving both CS and Math.
Requirement to be on campus: No
Supervisor: Sri AravindaKrishnan Thyagarajan , Pieter Roffelsen, Tomas Lasic Latimer
Eligibility: A WAM above 85 is required. Knowledge of modular arithmetic and statistics.
Project Description:
The dawn of the quantum era is near, yet much of our cybersecurity still rests on assumptions based on binary computing. In this interdisciplinary project, we search for new quantum safe algorithms by bringing ideas from the modern theory of integrable systems into the discrete world. To this end, we will stress test structures from integrable systems against both state-of-the-art classical attacks and the emerging threats posed by quantum computers. This will be an interdisciplinary project involving both CS and Math.
Requirement to be on campus: No
Supervisor: Dr Katy Gero
Eligibilty: Understanding of LLM training pipeline. Ideally experience with Python and the HuggingFace library.
Project Description:
Many organizations, partially driven by an interest in AI sovereignty, are developing nation-specific or region-focused language models. For example, LLM-jp in Japan, SEA-LION in Singapore, Apertus in Switzerland, and Latam-GPT in Latin America. In this project, you would map the landscape of these projects: how different groups approach the development of their model, comparing training datasets, and benchmarking these models for both general ability and cultural and linguistic fit.
Requirement to be on campus: Yes *dependent on government’s health advice
Supervisor: Dr. Katy Gero
Eligibility: Comfortable writing Python code and working with libraries like HuggingFace. An interest in the arts is necessary; this is not a project about optimizing benchmarks.
Project Description:
What are the best tools for creative practitioners interested in training tiny language models for artistic purposes? Test out different methods for training or fine-tuning tiny language models (<1B parameters) on small datasets (<50MB) and compare these to older methods like markov chains and LSTMs. Perform either an automatic evaluation (e.g. stylistic fidelity to training data) or a small user study in which artists compare the outputs of different models.
Requirement to be on campus: Yes, *dependent on government’s health advice
Last updated 12 September 2026.
