STEM Career Sprint 2026

From papers to products

Career Roadmap
for AI Researchers

Don't get swept away by the AI wave —
become the surfer.

Speaker

Yunsung Lee
WoRV @ Maum.ai

Venue

SUNY Korea
April 4, 2026

Who Am I?

10+ top-tier papers 1,850+ citations double-digit h-index

Opening roadmap

The wave is here.

The AI shift is already changing how people build, learn, and get hired. Today is about how to ride it on purpose.

Today's route

1

See the Wave

AI landscape, salary reality, and why the old playbook broke.

2

My First Wave

What shipping for 5M users at Wrtn taught me.

3

The Next Wave

Physical AI, research bets, and why robotics matters now.

4

Your Surfboard

A blunt career guide you can act on this week.

The fastest growth in history

200M

ChatGPT reached 200 million users in two years.
Now software developers sit at 8–9/10 on AI job exposure.

The surprising part is not that AI is coming. It is that the most screen-native, high-paying work is now directly in the blast radius.

Karpathy's warning

8–9 / 10

Developers rank near the top of AI exposure because their work already happens in language and interfaces.

Why it matters

2 years 200M users job map changed

Know Your Wave

AI Researcher

$160K - $400K TC

Publish papers, advance SOTA.
Deep math & experimentation.

AI Engineer

$180K - $320K TC

Build production ML systems.
High demand, model serving, infra.

Data Scientist

$130K - $250K TC

Analyze, model, interpret data.
Bridges research and product.

AI Product Manager

$150K - $280K TC

Define what to build and why.
Rarest combo of user & tech.

Salary reality check

The U.S. pays for AI already. Korea mostly hasn't — yet.

That mismatch is exactly why the window matters: the Korean premium is still small, but the capability demand is moving much faster than the salary signal.

United States

AI engineers already command a visible compensation premium.

Korea

Only a modest gap today, which means the upside is more about timing than current salary.

United States — total compensation

$245K

AI
Engineer

$180K

Non-AI
Engineer

+36% AI premium

Korea — salary gap

₩95M

AI
Engineer

₩90M

Non-AI
Engineer

Only 6% gap in Korea

50%

of AI positions remain unfilled

Companies can't find qualified candidates despite waves of applicants.

WHY?

6.1%

CS grad unemployment rate

Degree holders can't land jobs. Skills mismatch — market wants AI-augmented builders.

Anthropic  Mar 2026

75% of programming tasks are coverable by AI  ·  14% hiring slowdown for traditional coders. The old playbook is broken.

What students think matters

The old playbook.

• GPA / Honors
• Degree Prestige
• Certs & Bootcamps
• Traditional Resume

These signals are not meaningless.
They just no longer win by themselves.

VS

What recruiters actually want

The 2026 reality.

• Portfolio 38%
• Internships 35%
• Public Code 34%
• Credentials 4%

The strongest signal is still simple:
can you point to real things you built?

The ChatGPT moment for robotics

When machines begin to understand, reason, and act in the real world — that's Physical AI.

Jensen Huang · CES 2026

Capital

$7.2B

Robotics funding in 2025 — up 2.3× from 2023.

Units

13K–18K

Humanoids shipped in the first meaningful commercial year.

Korea signal

60%

Share of CES 2026 Innovation Awards tied to Korea.

NVIDIA Tesla Figure Boston Dynamics 1X Unitree

National AI Budget 2026

₩10.1T

206% increase YoY.
Your home advantage.

▲ Korea AI Pivot

Samsung Electronics

₩300T+

SK Group

₩247T

Hyundai Motor

₩156T

The 12-month paradigm shift

From vibe coding to agentic engineering.

A year ago the story was “let AI write it.” By 2026 the job is managing agents, systems, and feedback loops.

Feb 2025

“Vibe Coding”

Prompt the model, let it write, and optimize for speed and vibes. Great for momentum. Weak for sustained complexity.

→

12 months

Early 2026

“Agentic Engineering”

Break work into systems, orchestrate agents, verify outputs, and keep humans focused on judgment and design.

Jensen, Jul 2025: “Study physical sciences, not software engineering.” The winning stack is fundamentals + orchestration + domain expertise.

Two waves I caught

Transition

Two waves.
One lesson.

I caught a product wave and a research wave. Different contexts. Same truth: define the right problem first.

Why this section matters

Wave one

Wrtn — shipping fast, owning user outcomes, and learning what actually drives retention.

Wave two

WoRV — building systems that move from benchmarks to robots in the real world.

Throughline

Product judgment + research depth is the combination that compounds.

Part 2 · Product wave

Section divider

My first
wave

Riding the product wave at Wrtn —
how I helped build a 5M-user AI agent.

Why this section matters

Mode

Move fast, ship weekly, and feel user feedback immediately.

What to watch

Execution speed, product taste, and how small technical choices hit real retention.

Part 2 · My first wave

Korea's leading AI-native startup.

Your Own AI — the AI that's just yours. I joined when the wave was still forming and ended up owning product work end-to-end.

My role

AI Engineer — end-to-end product ownership

Monthly Active Users

5,000,000+

Monthly Active Users

One product. My team.

Autonomous AI agent — not just a chatbot.

Feature engineering

Memory made the assistant feel like it actually knew the user.

+12.7%

Week-1 retention

Userquery + history

→

Memory Storeretrieve long-term context

→

LLMreason with memory

→

Responsepersonalized output

Before

Generic assistant: every conversation starts from zero.

After

Remembering prior intent, preferences, and context makes the system feel like a product—not a demo.

Feature 01

Real-time Web Search RAG

User Query

→

LLM Router

→

Web Search

→

Synthesize

→

Live Answer

Feature 02

Voice Calls
GPT-4o Realtime API

API Launch

→

Integration

→

QA & Test

→

Production

API → Production

in 2 weeks

Product demo

This is what 5M users interact with every day.

Personalized conversations, proactive prompts, multimodal entry points — all inside one AI-native consumer product.

personalized chat proactive messaging search + voice consumer-scale UX

Life at a startup

The startup grind.

At a consumer AI startup, the work loop is brutally simple: ship, get surprised, recover fast, and ship again.

🚢

Mon

Ship

New feature goes live.

💥

Tue

It breaks

Users find the edge cases you missed.

🔧

Wed

Fix

Hotfix, triage, postmortem.

📋

Thu

Plan next

V2 is already being designed.

🚢

Fri

Ship again

And the loop starts over.

Honest Failure

What I
Got Wrong

"I optimized for technical elegance when users needed simplicity."

3

WEEKS

LOST

Reality Check

Skills That Actually Mattered

What I Thought

✗

PyTorch depth

✗

Paper count

✗

Algorithm elegance

✗

Complex architecture

What Actually Mattered

✓

Problem definition

✓

Speed of execution

✓

User empathy

✓

Shipping reliability

What You Should Build

→

Shipping ability

→

Communication

→

Curiosity

→

Bias for action

Path A

STARTUP

$90K – $220K

• Learn everything, own everything
• Ship fast, break things, fix faster
• High ownership, high stress
• Career breadth over depth
• Equity upside — if it hits

Path B

BIG TECH

$200K – $550K TC

• Specialize deep, move slow
• Process, roadmap, sign-off chains
• Stronger mentorship & resources
• Career depth over breadth
• Stable comp, brand name signal

Neither is "better." Pick the wave that matches your surfing style.

Product wave takeaway

The 5M users came from
problem definition, not prettier algorithms.

When the user problem is clear, the stack becomes simpler, faster, and more valuable.

Part 3 · Physical AI

Section divider

Catching the
next wave

From game data to real robots. From 29% to 97.3%.

Why this section matters

Why now

Physical AI is leaving the lab and becoming an economic platform shift.

What this section shows

How research moves from simulations, to benchmarks, to robots that actually work.

Maum.ai  ·  Part 3

World of Robotic Vision

General Autonomous Driving Agent — Navigation × Manipulation × Simulation

🧭

Navigation

VLA-based autonomous driving from sketch maps and natural language. SketchDrive deployed in orchards.

🤖

Manipulation

Robot arm control trained on game data. ICLR 2026 top 10%. Outperforms models 7× larger.

⚙️

Simulation

Sim2Real evaluation pipeline and agricultural robot benchmarks. GINT: 100+ farm deployments.

Head of Research

SketchDrive  ·  Navigation

Draw a sketch.
The robot drives.

Hand-drawn map → VLA → robot executes the route. Deployed in orchards. Production-ready.

2× faster inference
with InternVL

SketchDrive · Orchard Demo

Agricultural Robot Field Test

The Non-Linear Journey to 97.3%

29%

Sim-Only

→

🔄

Sim2Real

iterative refinement

→

97.3%

Final

D2E  ·  Data-to-Embodiment

Games
→ Robots

We trained robot manipulation using 1,300 hours of game data — collected in under 1 week by 1 person.

ICLR 2026  ·  Top 10%

1,300h

game data

96.6%

LIBERO success

beats 7×

larger models

D2E  ·  SO-101 Robot  ·  Side-by-Side

Watch the difference.

Baseline (FAIL)

VS

Ours (SUCCESS)

Open Source  ·  VLA Eval Harness & OWA

Open Source Accelerates Your Career

47×

Throughput Gain

VLA Eval Harness — 14 hours → 18 minutes across 13 benchmarks

400K+

Lines of Code

OWA toolkit — production-grade open-source evaluation infrastructure

13

Benchmarks

Unified harness covering all major VLA evaluation protocols

226 PRs  ·  1,960 commits  ·  All open source Open source = public résumé

CORE  ·  DGX H100 Cluster  ·  Infrastructure

Good infrastructure = good research

Before

Storage Bandwidth 10 MB/s

Capacity 28 TB

Cluster Legacy GPUs

→

After  (CORE)

Storage Bandwidth 1 GB/s  100×

Capacity 100 TB  3.5×

Cluster DGX H100

70,000+ jobs completed

"The bottleneck in research is rarely the idea. It's the infrastructure."

CostNav  ·  Research Meets Business

When research quantifies the real problem

Revenue

$1.5

Cost

$31.5

99.7% of costs from collision-related maintenance. Current collision rate: 54%. Need <5% for profitability.

GINT  ·  Research → Product Pipeline

Research That Ships

🔬

Lab

2024 Q3

SketchDrive core research begins.

🎬

Demo

2025 Q1–Q2

Sim & real-world demo. First field tests.

🌾

Pilot

2025 Q2–Q3

Sim2Real iteration. On-farm evaluation and tuning.

🚀

Commercial

2025 Q4

Commercial deploy complete. Product shipping.

"

Research that doesn't ship is just a hobby.

"

Researcher

rigor

novelty

first principles

patience

Engineer

speed

reliability

pragmatism

shipping

↓

OverlapResearch Engineer

Part 3 · Mindset

How to think like a research engineer.

🔬Define the research question and the success metric before you touch implementation.

⚡Build the simplest version that could work, then iterate with evidence.

📦A paper without code, or a model without deployment, is only half-finished.

The best researchers ship.

Research wave takeaway

Research still wins when it solves a real problem.

Benchmarks matter, but only when they change what becomes possible in the world outside the paper.

Part 4 · Career guide

Section divider

Your
surfboard

The blunt truth about what you need
to ride this wave — and what matters
far less than students think.

Why this section matters

Theme

Portfolio over posture. Shipping over signaling. Clarity over performative busyness.

Use this part for

Turning the talk into a practical career operating system.

Blunt

Your portfolio is your surfboard.

38%

Recruiters cite portfolio vs only 4% for credentials.

Source: Fortune recruiter survey, Dec 2025

GPA Transcript

3.87 / 4.0

Looks tidy. Says almost nothing about whether you can build.

VS

GitHub Profile

Active

Shows proof of motion, curiosity, and shipped work at a glance.

Part 4  ·  Practical

The Minimum
Surfboard

Three items. No exceptions. Start here before anything else.

✓

Active GitHub Profile

Green contribution squares matter. Commit daily, even small things.

✓

Tech Posts — 5+ Articles

Share what you build. Recruiters read it. Medium / dev.to / X / LinkedIn

✓

3 Deployed Projects with READMEs

Not just local. Live demo link. Documented. Anyone can run it in 5 minutes.

Part 4  ·  Blunt

Good vs. Bad Portfolio Project

Bad

Built a calculator app

No metrics Not deployed Generic

✗ Used Python and JavaScript

✗ Runs locally only

✗ No README or documentation

✗ Zero real-world impact shown

VS

Good

Built agent harness — 500+ tasks/week

Quantified Live demo Documented

✓ 500+ tasks/week orchestrated across agents

✓ Deployed with observability dashboard

✓ Full README with architecture diagram

→ Live Demo

Blunt

"If your GitHub only has homework, you're competing with everyone else who also did the homework."

Class assignments prove you followed instructions. Personal projects prove you can think. Recruiters know the difference.

student_2025

Empty — only homework

3 contributionsMostly assignments

builder_2025

Active — real projects

847 contributionsShips regularly

What to build this month

Pick one project. Ship it. Deploy it.

You do not need ten ideas. You need one concrete thing that looks real, runs publicly, and teaches you something hard.

Rule of thumb

A project should be understandable in 30 seconds and defensible in a 10-minute conversation.

Beginner

Personal dashboard

A polished utility with live data and a deployed URL.

API integrationdeployVercel / Netlify

Intermediate

Agent harness

A multi-agent orchestrator that plans, acts, and reports.

Claude / OpenAI Agents SDKLangGraphMCP tools

Advanced

Open source CLI

Solve a real pain point and publish something others can install.

real problemnpm / PyPIGitHub Actions CI

Part 4  ·  Not Optional

AI Tools You Must Know

41%

of all code is now
AI-generated or AI-assisted

C

Claude

Coding, writing, reasoning, long docs

Anthropic

G

ChatGPT

Brainstorm, interview prep, quick drafts

OpenAI

⌥

Cursor

AI-native code editor, tab completion

IDE

◈

Copilot

In-editor AI for GitHub repositories

GitHub

v0

v0

Instant UI components from prompts

Vercel

Part 4  ·  Inspiring

Build
10× Faster

10×

Students who use AI tools
ship 10× more projects

Same skill level. The only difference is the tools they use.

Before AI Tools

Plan &
Learn

Weeks

Write
Code

Months

Debug
& Fix

Weeks

Deploy
& Hope

Days

With AI Tools

Prompt
& Plan

Hours

AI Pair
Codes

Weekend

Deploy
Live

Hours

Iterate
Fast

Days

Transformation

The Resume

• 1
  Rewrite in STAR format
• 2
  Tailor for JD keywords
• 3
  Add real numbers

Preparation

The Interview

• 1
  Act as interviewer (Claude)
• 2
  Mock 20-min session
• 3
  Specific feedback loop

This is not cheating. This is the industry.

Interview framework

STAR makes vague work sound real.

Situation → Task → Action → Result. And the last word should almost always contain a number.

S

Situation

Set the context so the interviewer knows the stakes.

T

Task

Name the specific challenge you were responsible for.

A

Action

Explain what you actually did, not what the team did.

R

Result

Close with measurable impact and why it mattered.

Real example

Built a long-term memory + personalization RAG system at Wrtn, then tied the story to the business outcome instead of just the model details.

+12.7%

Interview Framework

Same work. Better framing.

Weak signal

“Worked on ML project for recommendation engine.”

Generic. No scope, no ownership, no metrics.

→

Translate
work into
impact

Strong signal

“Built a recommendation pipeline that improved CTR by 23% for 50K daily users.”

Ownership, scale, and business value — in one sentence.

Part 4  ·  Practical

Technical
Interview
Pyramid

Coding is table stakes. System design is the differentiator. Behavioral is what closes the offer.

Behavioral

STAR stories, cultural fit, conflict resolution

Closes offers

ML Theory

Fundamentals non-negotiable for AI roles

Differentiator

System Design

Scale, trade-offs, distributed systems — the senior filter

Senior filter

Coding (LeetCode / Algorithms)

Table stakes — everyone does this. Pass or you're out. Start here.

Required

Culture Fit Decoded

"Tell me about a conflict."

↓

Collaboration & Humility

Can you update your views with new evidence?

"What is your weakness?"

↓

Self-Awareness & Growth

Do you know yourself? Are you improving?

"Why work here specifically?"

↓

Genuine Interest

Did you research us, or just applying?

What everyone does

Skills Section

PythonTensorFlowPyTorch DockerAWSKubernetes SQLFastAPIGit RedisMLflowAirflow

Lists tools. Shows nothing. Recruiters skip this.

Blunt

→

Show what you BUILT.

What gets you hired

Experience Section

›

Deployed real-time inference pipeline on AWS serving 10K req/s with p99 latency <50ms

›

Built Kubernetes-based autoscaling system — reduced infra cost 40% during off-peak hours

Context + Action + Quantified Result. Every time.

Year-by-year roadmap

What to optimize each year.

Do not try to do all four years at once. Each stage has one main job. Nail that job, then move on.

Year 1

Explore

Master fundamentals in Python, math, and basic AI.

Get curious fast about what kind of wave you want to ride.

Year 2

Build

Start projects and use AI tools aggressively.

Ship your first real code instead of collecting tutorials.

Year 3

Ship & Intern

Target internships and public contributions.

Make your portfolio legible to a recruiter in 30 seconds.

Year 4

Optimize

Polish the strongest work and tailor for target roles.

Land the role with focused applications and proof of shipping.

Part 4 · Practical

Networking that actually works.

Your next job usually comes from someone you've helped, not someone you've merely introduced yourself to.

1

Open Source Contributions

PRs merged = proof of skill + visibility

Highest ROI

2

Conferences & Communities

PR12, TF Korea, NeurIPS, ICLR socials

High

3

LinkedIn — Writing & Engaging

Share your projects and learnings publicly

Medium

4

Career Fairs

Useful, but everyone is doing the same thing

Low

Best networking loop: ship something useful → help someone with it → stay in touch.

Blunt

80/20

20% of activities produce 80% of career results.

The 20% that actually matters

🗂

Portfolio Projects

Deployed, quantified, documented. The only thing that speaks without you present.

🤝

Networking — the right way

Open source, communities, conferences. Give value first.

🎯

Interview Prep

STAR stories, system design, coding. Deliberate practice, not cramming.

Cut the noise. Ignore the rest.

Part 5  ·  Action Items

Time to
Paddle Out

You've seen the wave. You've seen the surfboard.
Now get in the water.

01

Action Item · Do This Today

Install Claude Code
or Cursor TODAY

Yes, it costs money. Check student support programs. Anthropic and Cursor both have options. This is the most important investment you'll make this year.

Claude Code

Terminal AI agent

Cursor

AI-native VS Code fork

★  Student plans available — check before paying full price

02

Action Item · This Week

Start a project that
solves something annoying

Think of one daily frustration. Something that wastes your time. Build something that fixes it. Let Claude init the project for you.

Terminal

$ claude "init a project that solves X"

✓ Scaffolding project... creating README, setting up repo, drafting first feature...

→  Let AI help you start. You provide the problem. It provides the momentum.

Final step

Put an AI agent on your phone.

Make AI part of your default workflow — not a toy you only open when homework is due.

📅

Schedule planning

Turn vague to-dos into a real plan.

✉️

Email drafting

Write faster without sending careless messages.

💻

Code review

Check ideas, diffs, and bug hypotheses on the go.

📚

Instant learning

Ask while curiosity is still hot.

How can I help today?

Draft email to my professor

Done. Want a formal or friendly version?

Review my PR summary

Looks solid. One edge case still missing.

Part 5 · Recap

The wave is real —
and already hiring differently.

Before we recap my two waves, remember the setup: the market is moving fast, expectations are shifting, and the upside is unusually large for students who adapt early.

Compensation

$245K

Average AI engineer total compensation in the U.S.

Talent gap

50%

AI roles remain unfilled even while routine coding gets automated.

National bet

₩10.1T

Korea's 2026 AI budget — your home court advantage.

Part 5 · recap

What my two waves actually produced.

In both jobs, problem definition beat algorithm worship.

Product

5M+

Monthly active users

A real product people came back to every month.

Retention

+12.7%

Week-1 lift

Memory + personalization RAG changed the user experience.

Research

97.3%

GINT success

A benchmark jump strong enough to matter outside the lab.

Paper

ICLR

Top 10%

Research quality is still rewarded when it solves the right thing.

Part 5 — Recap

YOUR SURFBOARD

• 🏄

  Portfolio

  Build and deploy 3 real projects. Metrics matter. Deployment matters. GitHub commits matter.

• 🤖

  AI Tools

  Claude, Cursor, Copilot — use them daily. This is not optional. It is table stakes.

• 🌊

  Network

  Open source, communities, LinkedIn. Your next job comes from someone you've helped.

SCAN ME

LinkedIn

CONNECT

• linkedin.com/in/yunsung-lee-23a926150

  LinkedIn

• github.com/alohays

  GitHub

• alohays.github.io

  Personal Website

"Reach out. Visit the lab. Contribute to open source. I mean it."

The surfer's choice

Sam Altman, August 2025: “If I were graduating college right now, I'd feel like the luckiest kid in all of history.” Luck only matters if you act on it.

The wave doesn't care
if you're ready.
It's coming either way.

The only question is:
will you paddle out?

Write down one thing
you'll do this week.

Right now. On your phone.

I'll wait.

Thank you.