R.e.
Who am I

R.e.

Computer vision and applied AI, from research to something that actually runs. Pose estimation, medical imaging, behaviour analysis, LLM applications — and the engineering that keeps the claims checkable.

Computer vision · LLM applications Shanghai · Singapore English · 中文
R.e. R.e. Technical support · AI cloud, international
Now
Overseas AI deployments — scoping, POC, contract, production.
Focus
Model evaluation, computer vision, agent and LLM applications.
Regions
APAC · Europe · Middle East.
Method
Real calls, raw output, and a written record of what failed.
Illustrated portrait of R.e.
Illustration, not a photograph.
01
2023 · Independent research · EI conference

Predicting urban CO₂ emissions from mixed data

City emissions are driven by two very different systems — natural processes and human activity. I screened multi-source data with Pearson correlation, kept the drivers that actually moved the number, and modelled the trajectory with time-series and logistic fits to sketch scenario paths.

  • Pearson screening
  • Time series
  • Logistic model
EIICEMEE 2023
02
2023–2024 · Undergraduate thesis · Distinction

Joint injury risk from human pose estimation

A 24-keypoint detector plus a risk model that groups joint motion into five clusters and compares left-right strength asymmetry. Built on 16 subjects and 3,137 frames across 16 datasets; on drop jumps it located the at-risk knee with FNR 0.25 and FPR 0.167.

  • 24 keypoints
  • ResNet-34
  • Biomechanics
0.25FNR · 0.167 FPR
03
2024–2025 · A*STAR × NTU · Best report

Behaviour detection for a neurodegenerative disease model

Extending ResNet-50 pose estimation with depth cues, then feeding the motion graph into temporal models to predict disease risk in a primate model. The pipeline is the same shape as action recognition in the wild — detect, track, predict.

  • ResNet-50
  • Depth estimation
  • Temporal modelling
Top 1BMDSIS report
04
2022 · National electronic design contest

Vision-guided ward delivery robot

An on-board OpenCV pipeline for ward-number recognition — preprocessing, contour detection and template matching — with PID tracking and a wireless handshake that closed the drop-off and pick-up loop. Beijing second prize.

  • OpenCV
  • PID control
  • STM32
2ndBeijing
05
2022–2023 · Challenge Cup · Beijing third prize

Ultrasound scoring for placental accreta

U-Net with attention for segmentation across angles and lighting conditions, plus an expert-system layer that turns measurements into a clinical score. Third prize, Capital Challenge Cup.

  • U-Net
  • Attention
  • Expert system
The paper · pp. 131–147 Click a page to enlarge
3rdBeijing
Publication

Prediction of Carbon Emissions from Natural and Human Systems Based on Correlation Analysis and Logistic Models

R.e. · ICEMEE 2023
E3S Web of Conferences 406, 03030 (2023)
DOI 10.1051/e3sconf/202340603030

The paper separates the natural and human drivers of urban carbon emissions, ranks them with Pearson correlation, and fits time-series and logistic models to project how emissions evolve under different assumptions. It is a small paper, but it is the one I still use to explain one thing: pick the features before you pick the model.

Published under my legal name; this profile uses the R.e. handle. The publisher's PDF is the version of record.

Final year project

AI-assisted educational platform for Gastric Oncology

R.e. · Nanyang Technological University
Final Year Project (Practicum BS6211), School of Biological Sciences
A*STAR Bioinformatics Institute · 2025

This extends the A!HistoCloud pathology annotation platform into a teaching platform for gastric oncology — three roles (student, teacher, supervisor), course and homework flows, a shared community — with two AI layers: a local LLM that helps teachers draft courseware and summarise teaching records, and a multimodal model that reads annotations and slide images and explains them to learners. Built at A*STAR's Bioinformatics Institute as my NTU final-year project.

Download PDF 29 pages · 4.7 MB

Submitted under my legal name. The PDF is the version handed in to the school, unedited.

BMDSIS project

Detecting and predicting behaviors of neurodegenerative disease primate models

R.e. · Keng-Hwee Chiam
School of Biological Sciences, Nanyang Technological University
BMDSIS · 2025

The cage is in every frame, and that occlusion is what breaks pose estimation. This project compares two ways to remove it before tracking — a Canny + closing + mean-fill pipeline, and an AOT-GAN inpainting model — then measures the effect with DeepLabCut on six marmosets, using 3D depth estimation for the macro view and a per-frame Euclidean distance error for the micro view. AOT-GAN kept more detail and cut preprocessing from minutes to seconds per image; all six animals improved overall, Boba's error roughly halved, and the one regression (Horlick, first seconds) traced back to a hand touching the cage — which became a rule for auto-selecting frames.

Download PDF 11 pages · 1.9 MB

Submitted under my legal name. The PDF is the report as handed in.

That is the record so far. Papers are published under my legal name; R.e. is the handle I use here.