Napat Naowarangsee

LiDAR and AI researcher for patient monitoring

Date of birth: 2 July 2001

The face is drawn from my photograph. The other point clouds on this page are simulated. Each point takes its brightness from that photograph.

Napat’s portrait, drawn in points

I research patient monitoring with LiDAR and AI, and I develop two health platforms.

Research

I use LiDAR and AI to assess the body non-invasively. This room shows what that makes possible, and the research behind it.

Simulated LiDAR points of a patient room. The person lying on the bed is drawn in amber.

Reading posture from LiDAR points, without a camera

No camera. Only points.

Points captured0 / 16,384

No person found yet. Capture points up to the mark.

Move the mouse or tap the room to sweep the beam. When you stop, it sweeps by itself.

Real-time point cloud analysis and deep learning assess a patient's posture non-invasively and alert caregivers.

15.6 ms per frame on a Raspberry Pi 5

Master's thesis, KMUTT (ongoing) LiDAR-based Posture Monitoring Using AI

Posture classification on an edge device, toward pressure ulcer prevention

PublicationEdge-Embedded LiDAR Point-Cloud Posture Classification Toward Pressure Ulcer Prevention in Bedridden Patients

N. Naowarangsee, S. Boonprong, A. Khantachawana

BMEiCON 2026, Fukuoka, Japan

Published pp. 1–5

Keywords
LiDAR, point cloud, posture classification, PointNet, edge computing, pressure ulcer
Open the paper (DOI)

Lying on backSimulated result

15.6 msper frame on a Raspberry Pi 5

A drawing of the edge device: a Raspberry Pi 5 under its Active Cooler, joined by a USB 3 cable to a LiDAR L515 camera on a stand over a bed. The beam sweeps the bed, and a simulated point cloud of a person lying on their back appears, coloured by distance from the camera.

Estimating skin properties non-invasively, with an air jet and LiDAR

An air jet and LiDAR estimate skin properties non-invasively. The method analyses how the skin deforms, measured optically, for medical diagnostics.

< 10% Young's modulus error

Master's thesis, Kanazawa University (ongoing) Non-Invasive Skin Property Estimation Via Air-Jet and LiDAR

A mock circulation device for testing stent retrievers

A bench loop that stands in for the M1 artery of the brain. A heated chamber, a pump and a valve that opens every 0.75 s (80 beats a minute) drive a blood analogue through a vessel model. It was built to test a braided nickel-titanium stent retriever: at most four passes, scored by how much of the clot comes back.

50.45–121.36mmHg measured in the model (an artery: 80–120)

329.82ml/min average flow (an artery: 275.4)

The moving numbers replay the measured range.

Bachelor's project, KMUTT Design and Development of the Mock Circulation Device for Efficiency Testing of Stent Retriever

Education

  • Master's degree in Frontier Engineering (Human-Machine Symbiotic Systems Program)

    Kanazawa University

    Oct 2024 – present

  • Master's degree in Biomedical Engineering

    King Mongkut's University of Technology Thonburi (KMUTT)

    Jul 2023 – present

  • Bachelor's degree in Mechanical Engineering

    King Mongkut's University of Technology Thonburi (KMUTT)

    Jul 2019 – Jul 2023

Tools
Python, C++, PyTorch, TensorFlow, OpenCV, Open3D, Scikit-learn, Raspberry Pi, Intel RealSense SDK

Systems in use

Two health platforms I developed, and other software you can try.

CLICS

AI clinical reasoning platform

Sole developer

  • Scores physicians' clinical reasoning from their conversations with AI, for a licensing body's CME pilot.
  • Removes personal data before every call to Claude.
  • Handled 100 concurrent users with 0 errors in load tests.

Next.js, Supabase, Claude API

Jun 2026 – present

This clip has no sound.

The loop that plays at the CLICS booth: draw a made-up case, practise the diagnosis with AI, then see your result.

Active Ageing Index

National elderly-screening platform

Sole developer

  • In production. Field staff use it through LINE.
  • Combines 24 clinical scoring tools into one ageing index.
  • Hides small groups from dashboards (k-anonymity).

Next.js, Supabase, FastAPI, LINE API

Jun 2026 – present

Screenshot of the admin dashboard
Screenshot of the field staff home screen
Screenshot of the submission status screen
Screenshot of the form for one area
Screenshot of the grid for several areas
The platform’s menu in LINE
  1. The admin dashboard: the AAI score of every project, at three points in time.
  2. Field staff see their projects and this month’s submission status.
  3. Monthly data goes in one area at a time, as a grid, or from an Excel file.
  4. The menu in the LINE Official Account, the front door for field staff.

Screenshots show dummy data

ClaudePet

Open-source macOS app

  • Shows Claude Code usage and rate limits from local logs, with no network access.
  • Public on GitHub, with CI tests and releases up to v0.14.0.

Swift, SwiftUI

Jun 2026 – present

View on GitHub
Screenshot of the ClaudePet dashboard
Screenshot of the wide two-column widget
Screenshot of the tall single-column widget
The mascot in six moods
  1. The dashboard: the 5-hour and weekly limits, and tokens by model.
  2. The desktop widget, in its wide and its tall layout.
  3. The mascot’s mood follows how much of the limit is used.

Web applications

  • Workout tracker (PWA)

    A 3D muscle atlas of 64 muscles, linked to the exercise library.

    Open the demo
  • Nutrition planner (PWA) that keeps its data on the device

    Daily targets for energy, protein, carbs, fat and water.

    Open the demo

React, TypeScript, Vite, Vercel

May – Sep 2026

The app’s 3D figure pressing two dumbbells overhead
Screenshot of logging sets during a workout
Screenshot of the history summary: total volume, records and sets
The muscle body from the front and from the back
  • Primary
  • Secondary
  • Stabiliser
  1. Workout tracker: every set is logged by reps and weight.
  2. Its own screens: logging a session, and the total volume lifted.
  3. The 3D atlas shows the muscles of a Barbell Back Squat by role.
  4. Nutrition planner: daily targets for energy, protein, carbs, fat and water.

Sample data

Working with AI agents

I design agent workflows and context in Claude Code.

jjstack

jjstack is my own Claude Code plugin: skills with Python checkers, tested hooks and rules.

I test in a sandbox before deploy. CI runs the tests and the releases.

  1. 1grill
  2. 2spec
  3. 3tickets
  4. 4build
  5. 5gate
  6. 6merge
  7. 7lessons

7 stages

main session/jjstack:jjflow <idea>

  1. #1opushigh→ PR
  2. #2opusmedium→ PR
  3. #3sonnetmedium→ PR
  4. #4sonnetlow→ PR
gate/jjstack:doublej #<spec>
  • reviewer
  • verifier
  • fixer
  • Merge: auto
jjstack
0.1.134
skills
14
agents
3
lessons
37
hooks
6
  • allow-readonly-bash
  • load-gate
  • jjstack-main-guard

3 of 6 hooks check a command before it runs.

example run

  1. grillquestions until the task is clear
  2. specspec issue written, Merge: auto
  3. tickets4 tickets, each names its tier
  4. build4 sessions, each in its own worktree
  5. build4 pull requests open
  6. gategate-reviewer reads every diff
  7. gategate-verifier runs the checks: 3 PASS, 1 FAIL
  8. gategate-fixer fixes #4, round 1
  9. gate#4 checked again: PASS
  10. mergeMerge: auto, so 4 PRs are squash-merged
  11. lessonsthe lesson of the fix is kept for the next run
  1. The loop of jjstack: one plan, a session for each ticket, a gate on every PR.
  2. /jjstack:jjflow asks until the task is clear, then writes a spec and its tickets.
  3. Each ticket gets its own session, on the model and effort the ticket names.
  4. /jjstack:doublej reviews and verifies every PR, and fixes one that fails.
  5. PRs that pass merge as the spec says, and the lessons go back into the skills.

About 1,000 Claude Code sessions since May 2026

Tools
Claude Code, Codex CLI, MCP
Harness
skills, subagents, hooks, context engineering
APIs
apps built on developer API keys, such as the Claude API in CLICS

About1,000Claude Code sessions since May 2026

  1. implement #<ticket>
  2. planplan approved before any codegrilling
  3. exploresub-agents read in parallelsonnet medium
  4. worktreeits own branch and foldergit worktree
  5. testtests run before the PRnpm test
  6. browserheadless check, proof savedPlaywright
  7. reviewa second model reads the diffCodex CLI
  8. prpull request openedgh pr create
  9. 7 steps, from the plan to the pull request

One mark for each session. The bar shows the steps of one.

Awards

5 awards from invention and innovation events in Thailand, Taiwan and Germany

Silver Medal

iENA 2025, Nuremberg, Germany

Edge LiDAR system for pressure ulcer prevention

Nov 2025

Prominent Invention Special Award

Taiwan Prominent Inventor League, at iENA 2025

Edge LiDAR system for pressure ulcer prevention

Nov 2025

Bronze Medal

Taiwan Innotech Expo 2023, Taipei

Mock circulation device for testing stent retrievers

Oct 2023

Gold Medal

NRCT Higher Education Innovation Contest 2023, Thailand

Stent retriever test rig

Aug 2023

Outstanding Award

NRCT Innovation Project Proposal 2023

Stent retriever test rig

Aug 2023

Contact

To talk about a role or the research, email me.