Portfolio · Edition of October 2026
Research Fellow in Ophthalmology at Thammasat University. I build things that help people see: an AI that reads the retina, a test that reads tears, and simulators that let trainees practise before they touch a patient.
Diabetic retinopathy can be seen long before it costs anyone their sight, but only if someone looks. In rural Thailand there are few ophthalmologists to look. OphScan reads a fundus photograph, gives an international grade and shows the findings behind it, so a general practitioner can decide who needs referral.
Built on MedGemma-1.5 and fine-tuned with SFT and GRPO on LoRA/QLoRA adapters. In use for community screening at Khu Khot, Pathum Thani. A prospective study is now measuring how it changes GP decisions in primary care.
Ethics: Thammasat HREC MTU-EC-OO-0-038/69 (COA 071/2569) · Grant F2/2569
Dry eye is not one disease. Some eyes make too few tears, others lose them too fast, and many do both. The right treatment depends on which. I profiled tear peptides by MALDI-TOF and asked two questions. How repeatable is the measurement? And which peptide peaks separate the subtypes?
Concept Where it leads: a lateral-flow strip that reads three subtype markers at the point of care, in about ten minutes.
The candidate feature in Control, evaporative (EDE) and Mixed dry eye. The protein label is a tentative mass match.
3D PCA with confidence ellipsoids for each subtype.
Technical variance can be larger than the difference between people. Any tear biomarker study has to be designed around that.
M.Sc. thesis, Bioinnovation and Entrepreneurship, Chulabhorn International College of Medicine. Defended 22 September 2026.
Ethics: Thammasat HREC MTU-EC-OO-0-252/68 (COA 054/2569).
Neuro-ophthalmology is taught as diagrams that students memorise. I built a 3D model they can question instead: choose a lesion and see the field defect, the pupils and the eye movements it causes. Here is a small version of that idea. Pick a lesion.
Viewed from above · patient's left on your left
Left eye (OS)
Right eye (OD)
Fields drawn as the patient sees them · dark = lost · a teaching sketch, not for diagnosis
Afferent pathway, pupils (light, near and sympathetic) and ocular motility with all six muscles and CN III, IV and VI. Each finding comes with the anatomy behind it, drawn from BCSC Section 5 with page citations.
The 13 standard transthoracic echo views, cut live through one 3D heart. Slide and rotate the plane the way you would move the probe.
One segmented CT scan turned into 8 organ meshes, with axial, coronal and sagittal views linked to the 3D model.
Four procedures, each rebuilt as a browser simulator through a microscope view. Each one has an awake patient where the real case has one, complications that can happen, an outcome view and a debrief. A demo mode plays the whole case and hands control to you the moment you click.
Primary nasal pterygium, awake patient under local. Graft harvest and transfer (it can flip silently), glue or sutures, mitomycin C option. Outcome views at day 1, 1 month and 12 months.
Three-port PPV for a macula-off superotemporal detachment in a pseudophakic eye, with wide-angle viewing and a flat contact lens.
Pterional transsylvian approach through the operating microscope, including a demo of rescuing an intraoperative rupture.
Temporal clear-cornea phaco for an NS3 cataract: rhexis, chop, cortex, IOL injection and dialling. The demo plays the whole case in about 3 minutes 40 seconds.
pterygium-sim.pages.dev · runs in any modern browser · for education and demonstration, not clinical training
A digital peak expiratory flow meter, tested against spirometry and the Mini-Wright meter.
Schematic forced expiration · what PRAAN captures
Presented at APSR 2022, Seoul and EAACI 2023, Hamburg. Published in F1000Research 2024.
In a resuscitation you cannot stop to weigh a child. Laid beside the child, the tape reads off an estimated weight and the matching doses and equipment sizes, in colour bands by weight.
Bridging the Specialist Gap: Multimodal AI Decision-Support for DR Screening in Underserved Primary Care
PI, first author. Prospective study with GPs.
Incidental Calcification of the Trochlear Apparatus Mimicking an Orbital Foreign Body
Case report · presented at WOC2026
Interocular Asymmetry in Infantile FEVR
Case report, second author · poster at APVRS 2026
Weight-Estimating Tape for Paediatric Emergency Care
PI. Development and validation
4Es Non-Pharmacological Factors and Asthma Control
Formal analysis and writing
Technical Variance Exceeds Between-Person Variance in Tear Peptide Profiling of Dry Eye
First author of thirteen · from the M.Sc. thesis
Binary Referable Evaluation Conceals Half the Annotation Error in the APTOS 2019 DR Benchmark
First author. Translational Vision Science & Technology (Q1) · major revision invited
Silent Under-Referral in Diabetic Retinopathy Screening
First author, primary analyst. BMJ Open Ophthalmology (Q1)
Accuracy of Peak Expiratory Flow by PRAAN, Spirometry and the Mini-Wright Meter
First and corresponding author. F1000Research 2024;13:782 · DOI
Combined Hepatocellular-Cholangiocarcinoma and Its Mimickers
Annals of Diagnostic Pathology 2021;53:151770
Blue top edge = Q1 ophthalmology journal · bronze = published
OCTA Microvasculature in Obesity without Diabetes
SR and meta-analysis · first author · PROSPERO CRD420261469534
Stage 4 · extraction
Tear Fluid Proteins for DR Severity Staging
Systematic review · lead reviewer · PROSPERO CRD420261446084
Stage 4 · extraction
Epigenetic and Biological Aging Clocks in Ophthalmology
Scoping review · registrant, co-lead · OSF c9bvg
Stage 2 · abstract screening
APVRS 2026 · Gold Coast
Poster, paediatrics track: interocular asymmetry in infantile FEVR.
40th World Ophthalmology Congress · Prague
Presentation Pod, Ophthalmic Trauma. Also Session Chair, Trauma Pods Session #2.
INTARG 2026 · Katowice
Oral presentation of OphScan to the international jury.
EAACI Hybrid Congress 2023 · Hamburg
Oral presentation: PRAAN against spirometry. Research Travel Scholarship.
Thai Asthma Council Annual Meeting 2023
Oral presentation of the PRAAN study.
26th Congress of the APSR · Seoul
Development and validation of a digital peak expiratory flow device.
I guided the Department of Medical Services delegation, with the BKGI team, on its visit to BGI: a first-hand look at gene and cell therapy, and at how AI is changing the way healthcare is developed.
Before medicine there was debating, and it never stopped. It taught me to build an argument that survives a hostile room, which is what peer review asks for too. I debated for Thailand at the World Schools Championship, and I now coach a private team and adjudicate.
| Year | Tournament | Team | Speaker |
|---|---|---|---|
| 2018 | World Schools Debating Championship, Croatia and Slovenia | Competed | |
| 2016 | World Schools Debating Championship, Stuttgart | Partial Octo-finalist | |
| 2025 | Anglo Debate Blend | Grand Champion | Overall 2nd Best Speaker |
| 2024 | Kev's Academy Power Up BP (invitational) | Grand Champion | |
| 2024 | ICUT, International Christian University, Tokyo | Competed | |
| 2019 | 14th EU–Thailand National Intervarsity Debate Championship | Grand Finalist | Best University Speaker |
| 2017 | 2nd PDA High School Parliamentary Debate World Congress, Osaka | Champion, Native English Category | |
| 2017 | 3rd Asia World Schools Debating Championship | Semi-finalist | |
| 2017 | 2nd Assumption Thailand Debate Open | High-School Finalist | 5th Best Speaker |
| 2016 | 4th Thailand World Schools Debate Championship | Grand Finalist | 2nd Best Speaker |
| 2015 | 10th EU–Thailand National Intervarsity Debate Championship | Grand Finalist | 4th Best Speaker |
| 2015 | Forestry Debate, Food and Agriculture Organization of the UN | First Place | |
| 2015 | Heart of Europe, Czech Republic | Quarter-finalist |
Adjudication Core (invited), Thai National Medical Debate Championship 2024 and 2025 · Guest speaker for the Thammasat and Chula debate societies, 2023
Fine-tuning medical vision–language models with SFT and GRPO on LoRA/QLoRA. Shipped as OphScan.
Multi-reader adjudicated reference standards, abstention analysis, DeLong AUC, McNemar, design-weighted estimators.
Annotation-error audits of public DR benchmarks, with the reference standard fixed before the deciding reads.
Tool-using LLM apps that keep calculation apart from reasoning, triage first, answer from grounded retrieval or refuse, and log every tool call.
The explorers and simulators in chapters III and IV were built with agentic coding tools.
Python, TypeScript and React, Three.js, Git, FastAPI, Streamlit, BGE-M3 retrieval.
LinkedInScan or tap
ORCIDScan or tap
Try a simulatorScan or tap
