Jason Chia-Sheng Lin

Doctoral Researcher, Artificial Intelligence in Medical Imaging / Signal Analysis Lab National Yang Ming Chiao Tung University

I build evidence-aware AI systems for high-stakes work. My work brings together speech intelligence, clinical workflow support, agent governance, cybersecurity, LLM/RAG pipelines, and investigation-informed thinking to design systems that remain useful when evidence, review, reliability, and deployment constraints matter.

Professional Profile

Area	Details
Current role	Doctoral Researcher
Lab	Artificial Intelligence in Medical Imaging / Signal Analysis Lab
Institution	National Yang Ming Chiao Tung University
Location	Taiwan
Website	jasonln0711.github.io
LinkedIn	jason-lin-1a648813b
Email	cre062400@gmail.com

Summary

• I am a doctoral researcher in the Artificial Intelligence in Medical Imaging / Signal Analysis Lab at NYCU, where I work across trustworthy AI, AI Software as a Medical Device (SaMD), speech and language pipelines, and security-aware evaluation.
• Before doctoral research, I worked in cybercrime investigation. That background continues to shape how I think about evidence, adversarial behavior, failure analysis, and the difference between a strong model demo and a system that can actually be trusted in practice.
• My current work focuses on evidence-aware AI systems, medical cybersecurity governance, ASR + LLM + RAG workflows, runtime governance, and deployable AI for high-stakes environments.
• I am especially interested in collaborations that value technical depth, careful evaluation, human review, and realistic deployment conditions.

Core Focus Areas

Evidence-Aware AI Systems

Designing AI systems where reliability, evaluation, human review, and traceability are built into the architecture rather than treated as afterthoughts.

Speech, Language, and Evidence Pipelines

Building ASR + LLM + RAG workflows for conversational analysis, structured extraction, and evidence-aware reasoning over long-form audio and transcripts.

Security and High-Stakes Deployment

Studying privacy, leakage, adversarial risk, and governance constraints that shape AI systems used in regulated or security-sensitive environments.

Medical Cybersecurity and AI Governance

Current work includes cybersecurity risk management and governance for AI software medical devices, with attention to threat modeling, vulnerability and attack-surface analysis, Zero Trust implementation, and the interpretation of regulatory frameworks such as the U.S. FDA and Taiwan TFDA. I am particularly interested in translating standards and guidance into structured, verifiable engineering processes for cross-disciplinary teams.

Selected Projects

UroPrevisit Navigator

A synthetic urology previsit workflow for adaptive governed questions, missing-field repair, clinician-review summaries, and PSA / CRM-ready proposal support.

AI Triage Kiosk Demo

A synthetic vital-aware kiosk intake demo with governed English follow-up questions and staff-review summary output for a narrow market-demo scope.

Jarvis Voice Sight

A realtime voice-agent prototype with always-listening interaction, VAD, barge-in, turn isolation, sentence-level streaming TTS, and Ollama / vLLM runtime support.

Speech Evidence Intelligence Pipeline

An evidence-aware ASR + LLM workflow for turning long-form conversational audio into structured, reviewable outputs for high-stakes analysis. The pipeline emphasizes traceability between generated outputs and source transcript evidence.

Fraud Conversation Analysis with RAG

A retrieval-augmented workflow for analyzing fraud-related conversations while keeping language-model outputs grounded in transcript evidence. The project focuses on investigator-friendly review, grounding quality, and hallucination control.

Federated Learning Leakage Study

An AI security study centered on leakage risk and privacy trade-offs in federated learning for sensitive collaborative training settings. The work compares realistic threat models instead of treating federated learning as privacy-safe by default.

Selected Writing

• Recent Work, April-June 2026: Evidence-Aware AI Systems in Practice
• From Cybercrime Investigation to Trustworthy AI
• Designing Speech Evidence Pipelines with ASR and LLMs

Selected Teaching

• Teaching
• Enterprise AI Architecture Sprint
• Day 1: AI Gateway Architecture Evidence

Speaking

CYBERSEC 2026 Official Session

Title: AI 軟體醫材的資安實戰：從美國 FDA 524B 規範到 Threat Modeling 與 Patch SLA 的完整落地

• Event: CYBERSEC 2026
• Track: Medical Cybersecurity Forum
• Format: Breakout Session
• Schedule: May 6, 2026, 16:15-16:45
• Venue: Taipei Nangang Exhibition Center Hall 2, 4F Conference Room 4A
• Speaker page: CYBERSEC 2026 Speaker Profile
• Talks page: jasonln0711.github.io/talks
• Teaching page: jasonln0711.github.io/teaching

This session focuses on cybersecurity design for AI software medical devices, using FDA 524B as a practical anchor for threat modeling, SBOM, Zero Trust design, and auditable risk governance in heavily regulated environments.

CISC 2025 English Conference Papers

1. Evolution and Defense Challenges of Ransomware-as-a-Service in the AI Era
Technical and strategic analysis using Medusa and CrazyHunter as a case study.

• Event: Cryptology and Information Security Conference 2025 (CISC 2025)
• Schedule: May 28-29, 2025
• Venue: Feng Chia University
• Format: Conference Paper, English
• Conference site: CISC 2025

This paper analyzes how AI-era RaaS operations evolve through BYOVD, LOTL, covert C2, and adaptive tradecraft, then connects those threats to a ZTAID-grounded zero-trust defense strategy for practical containment and response.

2. Integration of Threat Pulse Modeling into the ZTAID Zero Trust Maturity Assessment Model
An analytical framework for continuous intelligence-driven assessment.

• Event: Cryptology and Information Security Conference 2025 (CISC 2025)
• Schedule: May 28-29, 2025
• Venue: Feng Chia University
• Format: Conference Paper, English
• Conference site: CISC 2025

This paper proposes Threat Pulse Modeling (TPM) as a way to transform live cyber threat intelligence into ZTAID pillar-level maturity signals, combining pulse-event mapping, severity triage, and time-series forecasting to accelerate the intelligence-to-assessment-to-response loop.

Experience Snapshot

Current

Doctoral Researcher, NYCU Artificial Intelligence in Medical Imaging / Signal Analysis Lab
Researching trustworthy AI systems, medical cybersecurity, speech intelligence, grounded LLM workflows, and security-aware evaluation for real-world deployment.

Previous

Cybercrime Investigation
Worked on digital evidence, online fraud analysis, OSINT, and operational reasoning in high-stakes investigative settings.

Ongoing

Research and Technical Communication
Developing case studies, technical writing, and speaking material around trustworthy AI, speech systems, and deployment risk.

Toolkit

ML / LLM Systems

PyTorch, Transformers, Whisper, LLM Pipelines, RAG Systems

Speech / Language

ASR, Speech Intelligence, Transcript Processing, Evidence Extraction, Conversation Analysis

Security / Operations

Cybersecurity, Digital Forensics, OSINT, Fraud Analysis, Federated Learning Security

Research / Evaluation

Experiment Design, Evaluation Frameworks, Reproducible Workflows, Python, GitHub Actions

Collaboration

I welcome thoughtful conversations around research collaboration, trustworthy AI, speech and language systems, and AI deployment in security-sensitive or regulated environments.

• Recent work: jasonln0711.github.io/now
• Research: jasonln0711.github.io/research
• Systems: jasonln0711.github.io/projects
• Contact: jasonln0711.github.io/contact

About This Repository

This repository contains the SvelteKit static source for my personal website, research pages, and MDX blog. The current public direction is a minimal academic/personal site: plain navigation, a readable homepage, chronological writing, and long-form article pages for trustworthy AI systems, speech intelligence, cybersecurity, regulated AI deployment, talks, and project notes.

The site keeps the older research/project/talk URLs available, while /blog/ is the canonical writing surface and /writing/ remains as a compatibility alias. Teaching accelerators are published under /teaching/, with the AI Systems Engineering Handbook repo kept as the canonical source for worksheets, instructor guides, rubrics, reference answers, and later day packages.

Current Implementation

• Framework: SvelteKit with @sveltejs/adapter-static.
• Styling: native CSS tokens and route-level Svelte styles.
• Blog content: top-level src/content/blog/*.mdx files compiled with mdsvex.
• Content layer: src/lib/content/site.ts for profile, recent-work, research, systems, teaching, and contact data, plus src/lib/content/blog.ts for blog metadata.
• Article UI: semantic long-form layout with generated h2/h3 heading IDs and a hydrated desktop table of contents.
• Static output: build/.
• GitHub Pages support: public/.nojekyll, robots.txt, /sitemap.xml, /rss.xml, and .github/workflows/deploy.yml uploading build/.

Key routes:

• /
• /now/
• /design/
• /audiences/
• /research/ and /research/[slug]/
• /projects/ and /projects/[slug]/
• /teaching/, /teaching/[accelerator]/, and /teaching/[accelerator]/[day]/
• /talks/ and /talks/regulated-ai-cybersecurity/
• /blog/ and /blog/[slug]/
• /writing/ and /writing/[slug]/ as compatibility aliases
• /about/
• /contact/
• /resume/
• /links/
• /zh-tw/
• /zh-tw/contact/

Design And Content Records

Historical v3.0 design records:

• docs/site-v3-clean-slate-rebuild-plan.zh-tw.md
• docs/site-v3-design-source.zh-tw.md
• docs/site-v3-content-inventory.zh-tw.md

The older section concept images used by /design/ live in:

public/design/v3-concepts/

Earlier v2.0 records remain useful as historical design context:

• docs/site-v2-information-architecture.zh-tw.md
• docs/site-v2-warm-evidence-studio-interface.zh-tw.md

Commands

npm run check
npm run build
npm run preview

npm run check runs SvelteKit sync plus svelte-check. npm run build writes the static site to build/.

Add A Blog Post

Create a top-level .mdx file in src/content/blog/, for example src/content/blog/new-note.mdx.

Use frontmatter like this:

---
title: New note title
description: One sentence summary for the index, RSS, and metadata.
pubDate: 2026-06-14
tags:
  - Trustworthy AI
  - Systems
category: essay
draft: false
featured: false
ogImage: /og/default.png
---

The route defaults to the filename slug, so the example publishes at /blog/new-note/. Add routeSlug: custom-slug only when the public URL should differ from the filename. Use normal Markdown/MDX headings; h2 and h3 headings become the article table of contents.

Add An Accelerator Day

Keep canonical course material in the sibling ai-systems-engineering-handbook repo. For the website, add a compact public-facing day entry in src/lib/content/site.ts under teachingAccelerators.

For a future Day 2 or Day 3:

1. Add or update the days object with a stable slug, title, sourcePath, sourceHref, public summary, learning outcomes, deliverables, sections, lifecycle, vocabulary, risk controls, worksheet prompts, and next gate.
2. Set published: true and give it a stable href such as /teaching/enterprise-ai-architecture-sprint/day-02-agent-governance/.
3. routePaths and the dynamic SvelteKit routes will include the page automatically.
4. Keep instructor-only reference answers, grading details, and private planning context in the handbook repo, not on the public site.

Validation Snapshot

Last verified after adding the teaching accelerator routes:

• npm run check
• npm run build
• Production preview smoke checks for /teaching/, /teaching/enterprise-ai-architecture-sprint/, /teaching/enterprise-ai-architecture-sprint/day-01-ai-gateway/, and /sitemap.xml.