How can researchers work effectively with an AI coding agent—and critically evaluate its results? This interactive Hackathon uses epidemic forecasting to explore the research workflow: framing questions, giving instructions, reviewing code, challenging assumptions, and evaluating results.
Starting with eight observations recorded at two-week intervals from a real measles outbreak, participants will work with Nguyen Tran using OpenAI Codex to explore the data, identify missing information, consider up to three candidate models, and forecast the remaining eight observations. Human scientific judgment will guide each stage.
- Date: Tuesday, 10 November 2026
- Time: 05:00–07:00 UTC (two hours)
- Local start times: 12:00 Bangkok / 13:00 Perth / 15:00 Brisbane / 10:00 Islamabad / 06:00 Lagos / 08:00 Nairobi
- Instructor: Nguyen Tran
- Read the complete participant guide
- Download the clean starter ZIP
- Download the Markdown participant guide
- Browse the session materials on GitHub
Who this is for
All experience levels are welcome. No advanced programming skills or individual Codex access are required. The main workflow will be facilitator-led, so everyone can join the discussion and scientific review. Participants who wish to run the Codex workflow on their own computers will need their own Codex access and the setup described below.
What you will learn
By the end of the session, you will be able to:
- Describe a practical workflow for collaborating with an AI coding agent in scientific research.
- Understand how a coding-agent workflow differs from a conventional chatbot conversation.
- Identify where human guidance, code review, and scientific judgment are essential.
- Explain the difference between fitting observed data and forecasting unseen observations.
- Recognise how limited data and analytical assumptions affect predictions and uncertainty.
- Explain why evaluation criteria should be agreed before withheld observations are revealed.
- Recognise the importance of reproducibility, held-out validation, and preventing data leakage.
- Critically assess AI-generated results, distinguishing evidence from assumptions and hindsight.
Before the session
Bring a computer with a reliable internet connection and access to the free version of ChatGPT for accompanying activities. The opening exercise uses CSV upload and data analysis; if those capabilities are unavailable in your account, follow the demonstration.
To run the full workflow yourself, also prepare:
- Codex access, signed in and able to open a local project folder;
- a working Python environment with NumPy, pandas, SciPy, and Matplotlib available to Codex; and
- Git, optionally, to record your own forecast checkpoint.
Download and extract the starter ZIP into a new folder. It contains only WORKSHOP_BRIEF.md and data/observed.csv. Keep a separate copy for the opening plotting demonstration, then use a fresh starter copy for the formal challenge. Open that extracted folder in Codex rather than the whole Hackathon repository, which contains materials from other sessions. You do not need to install Quarto to follow the lab.
Session outline
The two-hour session moves through six stages:
- Compare workflows: use ChatGPT and Codex to explore the same data, examining context, tools, and saved project files.
- Establish project rules: document known evidence, missing information, and the boundary around withheld observations; check that a fresh session can recover the saved context.
- Review a scientific plan: select at most three candidate models with explicit assumptions, uncertainty, and diagnostic checks.
- Implement and inspect: watch the agent write and run code, resolve errors, produce figures, and examine sensitivity to an important assumption.
- Freeze and evaluate: agree on evaluation criteria and save the forecasts and uncertainty estimates before the facilitator releases biweeks 9–16. Compare the frozen predictions with those observations.
- Conduct a scientific postmortem: discuss what worked, what failed, and how assumptions shaped the results, preserving the distinction between prior reasoning and hindsight.
The target is reported cases, not all infections. Population size, vaccination coverage, reporting completeness, and other community metadata are not supplied. Model choice remains open; participants are not required to use an SIR model. A convincing fit to the rising segment does not establish that a model will forecast the future well.
Use only the supplied observations before the reveal. Do not search for the original outbreak or unreleased outcomes. The facilitator will supply the withheld observations after forecasts have been frozen.
About the instructor
Dr. Tran Dang Nguyen is the founder of the Center for Computational Epidemiology and Infectious Diseases (CEID) in Vietnam, established following his tenure as a Research Assistant Professor at Temple University and Penn State University. He earned his PhD from The Open University, UK, in 2016, focusing on individual-based simulations to combat the spread of antimalarial drug resistance.
During his doctoral and postdoctoral work, Dr. Nguyen developed and validated a microsimulation model for malaria, used to evaluate strategies including multiple first-line therapies, triple artemisinin-based combination therapies (TACTs), and mass drug administration (MDA). His research has informed policy decisions in Uganda, Rwanda, and Tanzania.
The participant guide and starter materials were supplied by Nguyen Tran for this session.