The studio is designed for learners ready to move beyond a closed assignment. Independence is introduced in stages: first by reproducing evidence, then by changing one modelling choice, and finally by defending a question and evaluation plan of their own.
Mentored research in practice — AI4Nature@AVSS 2026
I mentored the student team behind From Pixels to Patrols: A Calibration-Aware Sensor Fusion Pipeline for Camera-Trap-Driven Anti-Poaching Resource Allocation. The official workshop programme lists the work as Paper ID 7 in Oral Session 1A on 31 August 2026; I am also listed as a co-author. View the official programme →
Mentored research in practice — ADMA 2026
I mentored the student research behind When Solar Power Changes Fast: A Multi-Site Audit of Prediction-Interval Reliability under Ramp Events and Sensor Degradation. The authors, in formal order, are Hong U Lo, Zibo Gao, Peng Chi Lam, and Sok Kin Cheng, all affiliated with St. Joseph Diocesan College (The Fifth School), Macao SAR, China. I am also the corresponding author. The work was accepted from the Research Track as a Short Paper on 3 September 2026; the camera-ready was submitted on 9 September 2026. The conference is scheduled for 13–15 November 2026 in Hong Kong; proceedings details remain to be confirmed. Research article → · Project and results → · Official conference →
Entry evidence
A learner is ready to begin when they can:
- explain a baseline model in their own words;
- read a labelled result table or figure critically;
- modify and rerun a small computational experiment;
- identify at least one assumption that affects interpretation;
- keep a dated record of decisions and results.
Advanced mathematics is helpful for some projects, but disciplined documentation is a more important entry condition.
Five research gates
Gate 1 — Reproduce
The learner regenerates one approved figure or table and explains the calculation behind it. The goal is not to copy code but to establish that the evidence chain is understood.
Gate 2 — Perturb
One parameter, input, boundary condition, or modelling assumption is changed. The learner predicts the direction of the effect before running the experiment and records whether the prediction was supported.
Gate 3 — Validate
The learner designs a check that could expose a weakness: a limiting case, alternative baseline, held-out scenario, numerical refinement, or sensitivity analysis.
Gate 4 — Formulate
Only after the first three gates does the learner propose an independent question. A short protocol defines the primary metric, comparison, evidence threshold, and scope before additional computation begins.
Gate 5 — Communicate
The final output contains a research question, model boundary, methods, claim-evidence map, limitations, and reproduction record. Public prose is written manually from reviewed results; private notebooks and raw experiment notes remain private.
Mentoring structure
Meetings are organised around decisions rather than progress percentages. A compact agenda asks:
- What changed since the last review?
- Which saved result supports the current claim?
- What is the strongest competing explanation?
- What is the next decision, and what evidence is needed for it?
The learner maintains ownership of the model while the mentor challenges scope, validation, and communication.
Research record
Each project keeps:
- a question and scope note;
- an assumption and data-provenance register;
- dated experiment entries;
- configuration files or notebook parameters;
- machine-readable result tables;
- a claim-evidence map;
- approved figures separated from temporary plots;
- a limitations and next-evidence note.
Suitable starting cases
Student-suitable entries can be explored through the project index. Public research notes such as When the Timetable Lies to the Ventilation System offer a bounded question, explicit counterevidence, and figures that can be critiqued without exposing the private computational workspace.
Completion standard
Completion does not require a positive result. It requires a transparent question, a coherent model, an appropriate comparison, inspectable evidence, and a conclusion that stops where the evidence stops.