students satisfiedWith my teaching, in Deakin's eVALUate survey.
807
GenAI policies analysedACPHIS-funded study, two papers accepted in 2026.
5
units using thucn.comIn 2026. The learning site I built.
“Python is not my favourite thing but it was easy to follow along and understand not just what he was doing, but also how and why so I can do it myself.”
Student, 2023
“He personally created a website for the content each week… probably would not have survived without this website”
Student, 2026
Teaching
GenAI can finish a prescribed analysis in minutes. I teach students to direct it. They frame the problem, choose the method, check the result and defend the decision.
Units
Taught ten units, including machine learning, business analytics and AI for business. Chaired two of them and coordinated a teaching team of six in a unit of 405 students.
Evaluations
95% of student responses were satisfied with my teaching (294 of 311, 2023 to 2026).
Design
Co-developed two new machine learning units. Set assignments on real datasets, such as NAPLAN results and crash hotspots, with industry speakers briefing the class.
Sharing practice
Presented at the Deakin Teaching and Learning Conference in 2025 and 2026, and co-ran a staff workshop on AI in teaching and research.
How students direct GenAI
One class in my introductory machine learning unit. Groups of four decide where to set a churn model's threshold and present their choice before the class ends. I use a Socratic approach. My questions guide each group, and we discuss the critical points so students learn to ask those questions themselves. The GenAI replies are illustrative. The costs and figures are the ones groups get from the case documents and data.
Step 1 of 5The instructor sets the steps and asks the questions
Instructor
Before anyone opens a GenAI tool, each group of four gets two things from me.
The analytical steps. Frame the decision, prepare the data, build the models, evaluate them, set the threshold, advise the client.
A template for the decision. Options considered, evidence, choice, and what would change it.
From there I guide with questions, and we discuss the critical points as they come up. At the threshold step the first question is “What does each mistake cost, and who pays for it?”
Students
The group reads the case documents.
GenAI
Not in use yet.
Step 2 of 5The group finds the costs and benefits and directs GenAI
Instructor
“Which figures in the case do you need before you ask GenAI anything?”
Students
The group searches the case documents for what each outcome costs and who answers for it, then writes its prompt from what it found.
“Southern Cross Telecom sends a $100 credit as a retention offer, paid whether or not it works. A customer who stays brings $300 in profit over the next year. Tom in Retention answers for the churners we miss. Lucas in Marketing answers for the offers we waste. Give me three ways to set the decision threshold, with pros and cons.”
The group tests each option against the unit material and the case. It rules out the first two, because neither uses what a mistake costs, and asks GenAI to apply the third.
GenAIIllustrative
Keep 0.50. It is simple and the most accurate setting, at 80%. It misses 160 of the 374 churners.
Maximise the F1 score. It balances missed churners against wasted offers. It ignores what each one costs.
Maximise net benefit. It uses your $100 and $300. It is only as good as those two figures.
Applied to the 1,407 test customers, net benefit is highest at a threshold of 0.37, at $32,000.
Step 3 of 5The group checks the result against the data
Instructor
“How do you know 0.37 is right?”
Students
The group reruns 0.37 on the test customers and moves the threshold itself. Try it below.
GenAI's figure holds. The check also shows something GenAI did not say. Between 0.33 and 0.50 net benefit stays within $1,300 of its peak, while missed churners and wasted offers trade places.
F1 peaks at 0.37 as well, so at these costs the two methods agree. Press “Replacement cost added”, where a customer who stays is worth $400, and the best threshold moves to 0.34. F1 does not move.
GenAI
Its answer is the thing being checked.
3. What the threshold gives
Predicted
Gets the offer
No offer
Actual
Would churn
True positive261Churners reached. Each one stays, worth $300, for a $100 offer.Benefit $78,300, cost $26,100
False negative113Churners missed, on Tom's account. $300 lost each, the same as with no offers.$33,900 lost
Would stay
False positive202Offers wasted, on Lucas's account. $100 each.Cost $20,200
True negative831Left alone. Nothing spent.$0
Click the chart to move the threshold.
At 0.37 the model reaches 261 of 374 churners and sends 202 offers to customers who would stay. Precision is 56%, recall 70% and F1 0.62. The benefit is $78,300 and the offers cost $46,300, so net benefit is $32,000, the highest of any threshold.
Net benefit is the benefit less the cost of the offers, measured against sending no offers. Sending none spends nothing and loses all 374 churners, worth $112,200.
Step 4 of 5The group questions GenAI and decides
Instructor
“What would have to be true for your threshold to be wrong?”
The template from step 1 holds the decision.
Students
“What did you assume to get 0.37?”
“What would change your answer?”
The group chooses 0.37. Compared with 0.50 it reaches 47 more churners for Tom, at the price of 85 more wasted offers for Lucas.
The choice carries a condition. If the first month shows that only about six in ten offers work, the threshold goes back to 0.50.
To see it, press “Six in ten accept”.
GenAIIllustrative
That every customer who gets the offer stays, and that each one brings $300.
The share of offers that work. If six in ten work, a churner you reach is worth $80, not $200. The best threshold rises to 0.50 and the net benefit falls to $5,420.
It makes no decision.
Step 5 of 5The group presents and the class questions it
Instructor
Each group presents before the class ends. The presentation is not marked. I ask every group the same four questions.
Which alternatives did you weigh?
How did you check the result against the data?
Which of GenAI's assumptions did you test?
How does your choice follow from what each mistake costs the two managers?
Then the class joins in. What if the offer costs more next year? Which customers were missing from the data?
Presenting is practice at explaining a decision to the people who must act on it.
Students
The group presents its recommendation and reads the same result twice, once for Tom and once for Lucas. Under questioning, it defends the choice or changes it.
GenAI
Not in the room.
The same pattern repeats at every analytical step, from framing the problem to the final advice.
What students say
Every small phrase below is a student's own, from anonymous evaluations of my teaching, 2023 to 2026. Unit names are hidden.
Click a trait to read its comments.Tap a trait to read its comments.
Now. Preparing week pages for a postgraduate unit on databases and business intelligence.
I study how universities govern GenAI. Early results show that policies on the same topic read alike, whichever university or country wrote them. A university can use that to audit its own.
807
policy documents
46
universities
GenAI policy documents from universities in Australia and Aotearoa New Zealand.
GenAI governance in universities
I led the first work package of Generative AI in IS Education and Research, collecting and analysing the policy documents. I am a Chief Investigator on this multi-university study, funded by the Australian Council of Professors and Heads of Information Systems (ACPHIS), 2025 to 2026.
It is my teaching problem seen from the institution's side.
2026
Fast AI, slow policy: Governing generative AI in higher education
Alam, S. L., Nguyen, L., Jha, M., Chen, J., Balapumi, R., Nguyen, T., and 15 others
Research-in-progress paper, Australasian Conference on Information Systems (ACIS 2026). Accepted. Preprint on request.
2026
Governing generative AI in higher education: Policy insights from an Australasian study
Alam, S. L., Nguyen, L., Nguyen, T., and 14 others
Short paper, International Conference on Information Systems Education and Research (ICISER 2026). Accepted. Preprint on request.
Emotion in tourism images
My longer programme builds AI systems that read and reshape the emotion in tourism images. My PhD, finished at Deakin in 2025, built one that estimates the emotion a destination photo evokes, then edits the photo to shift it.
CalmExcited
Drawn to feel Content
This drawing illustrates the idea and is not the system itself.
2024
A framework for destination uniqueness identification through concept modeling of tourist-generated travel photos
Nguyen, T., Vu, H. Q., Cybulski, J. L., and Nguyen, L.
Hawaii International Conference on System Sciences (HICSS 2024). HICSS proceedings.
2022
A design framework and AI system for affective destination image generation to influence tourists' emotional response
Nguyen, T., Nguyen, L., Vu, H. Q., and Cybulski, J.
Australasian Conference on Information Systems (ACIS 2022). ACIS proceedings.
Now. Writing a journal article on visual sentiment analysis for tourism marketing images.
I build the learning site, AI teaching agents and setup tools my classes run on.
thucn.com
The learning site I built and maintain, used by five units in 2026. Each week has its page, group activity and assignment guide, and students can leave anonymous feedback.
A Socratic tutor for a postgraduate machine learning unit, which questions students on the content and the assignment. I build agents like it on the Cogniti platform and am one of ten Cogniti Champions in the Faculty of Business and Law at Deakin.
Setup tools
One command installs the database and business intelligence software a class needs, with every download pinned to a version so each student gets the same install. Open the database setup or the business intelligence setup on GitHub.
Now. Extending thucn.com to more units and building agents for more of them.
Before my PhD I worked in AI and machine learning engineering at Ford Motor Company, as an intern from 2019 to 2020. For my PhD in Information Systems and Business Analytics at Deakin University, finished in 2025, I built an AI system that estimates and shifts the emotion in tourism photos. I have taught at Deakin since 2020, starting as a tutor during my PhD.
Get in touch
I am open to lecturer roles, staff training in analytics and AI, and research collaborations.