Guest Lecture Information Jungsuk Jay Lee (이정석) — ThroughLine · Head of AI Product & Advisor
Partnered with 1,600 crisis helplines across 176 countries, providing AI-powered mental health crisis classification (FirstStep) to global platforms including Google, OpenAI, Anthropic, Roblox, Discord, and YouTube.
F1 accuracy: 95.2% (all risk levels), 97.9% (all categories) Latency: 0.8s (p95) Cumulative users: 4M+, helpline contacts: 500K+ (130+ countries) Today's Class Structure Period 1+2 (100 min) — Guest Lecture by Jungsuk Jay Lee — The Reality and Technology of AI Crisis Support + Q&A Speaker intro: ThroughLine Head of AI Product, formerly Cookie Run (200M users), Naver, ZEPETO, Intel Problem definition: Psychological barriers (fear, stigma, mistrust) and fragmented crisis support ecosystem Real data: 90%+ of 4M users don't reach out even with free support; 50% are too scared to talk to a real person Structural limits of existing helplines: queue bottlenecks and one-size-fits-all processes ThroughLine's approach: AI classifies, assesses, and routes; humans focus on what only humans can do FirstStep system: Listen → Assess → Route Crisis classification: 5 categories (suicide, self-harm, eating disorder, online scams, bullying) × 5 risk levels (Critical–None) Technical challenge: accuracy vs. latency tradeoff — Rubric classifier (12s / higher accuracy) vs. Simple classifier (1.8s / lower accuracy) F1 Score explained: balancing Precision and Recall Solution 1: Supervised fine-tuning with clinician-labeled data → F1 from 82.1% to 87.8% Solution 2: Agentic Context Engineering — a self-improvement loop where the model refines its own prompts Final results: F1 95.2% (all risk levels), 97.9% (all categories), 0.8s latency — 2× more accurate, 2.3× faster Safe routing: tailored support by crisis level (self-help de-escalation → crisis counselor) Live demo + Q&A Period 3 (50 min) — Group Discussion Based on Guest Lecture Choose from 3 discussion topics for group discussion (20 min) Group presentations (5 min each) Key takeaways and wrap-up Period 3 Group Discussion Topics ThroughLine's FirstStep integrates with 1,600 helplines across 176 countries and performs real-time crisis classification on platforms including Google, OpenAI, and Anthropic. Behind the numbers — F1 95.2%, 0.8s latency — lie countless decisions at the intersection of technology and human life.
Topic 3: How Should AI Suicide Prevention Differ in the Korean Context? Cultural context
Korea's 1324 Suicide Prevention Helpline sends automated replies saying it 'cannot provide timely support' due to excess demand. Could the ThroughLine model structurally solve this problem in Korea? ThroughLine's 5 crisis categories (suicide, self-harm, eating disorder, online scams, bullying) were designed for English-speaking contexts. How should categories or risk levels change to reflect Korean language and cultural characteristics? If FirstStep were deployed on Korean platforms like Naver or Kakao, what design considerations arise between the Personal Information Protection Act and the Mental Health Welfare Act? Each team selects one of the three topics above, discusses for 20 minutes, then prepares a 5-minute presentation.
