Jason
KO

Visual lecture materials and agent automation

I’m Jason, an explorer at the frontier of knowledge.

Every day I look for new technology, try it myself, and test the ideas that come from it.

I majored in math and have been teaching for 10 years. Besides math, I now also teach machine learning, deep learning and AI.

News

  1. Joined an AICE education company; teaching ML and DL certification courses

  2. Bearable PoC completed (team lead)

  3. Two Apps in Toss mini apps launched

01

Lecture materials and teaching

2026.10 –AI education companyCourse materials, teachingAI certification courses at high schools and universities

Gradient descent scene from the 3D linear regression lecture. Over a purple loss surface, the update rule subtracts learning rate times gradient from the current W and b.
Gradient descent: updating W and b on the loss surface Run it (Korean narration)
  • Linear regression
  • MSE
  • Loss surface
  • Gradient descent
  • Voice narration
3D neural network lab. Nodes and links run from 8 inputs through three hidden layers to a predicted house price of 148,756 dollars.
Neural network: forward and backward pass
3D CNN lab. A 28 by 28 clothing image passes through convolution and pooling layers into 10 class probabilities.
CNN: filters and feature maps
3D random forest view. 32 trees stand in rows, each with its prediction path marked in blue.
Random forest: the average of 32 trees
  • 6 models
  • Trained on real data
  • Fashion-MNIST
  • Step-through calculation
Data type slide. A horizontal tree splits data into numeric, string, categorical and boolean types with example values.
Data type tree
Scaling slide. Scatter plots before and after normalization sit side by side, tracking how point B moves.
Before and after scaling
  • AICE Basic
  • For high school students and non-majors
  • 75 pages
  • Step-by-step animation
  • Apple Pencil annotation
A biological neural network and an artificial one drawn side by side in the same layout: one neuron and one node enlarged on top, many of them connected below.
Neural network: neuron to node Run (Korean)
A cat photo turns into 4×4 pixels and R, G, B number grids; 48 input values pass through hidden layers and come out as dog or cat.
Deep network: photo to class Run (Korean)
A 3D MSE loss surface where the current W, b point moves down to the next one, next to the prediction line and the b update calculation.
Gradient descent: formula and surface together Run (Korean)
  • Step-by-step playback
  • Real computed values
  • SVG
  • No libraries
  • Light and dark
02

AgentHub

Personal project2026.10 –In use

AgentsClaude CodeCodexCursorGrok CLIMCP, skills, rulesAgentHubOne Git repository22Skills17MCP tools12Notes4Rules3ModsDevicesMacBook AirMain workGit syncSamsung laptopAlways-on serverJob dispatch (SSH)DesktopGPUPower control (WOL)Dispatch order: Cursor → Grok → Codex → Claude (skips agents at 90% usage or more) · as of 2026.11

One repository holds the skills, tools and rules that Claude Code, Codex, Cursor and Grok all use. A change made on one device reaches the other devices and agents right away, and long jobs go to another device that sends back only the result.

A public copy of the flow diagram I use. Names, paths and databases are covered with black bars; the dots show which way data flows. Open full size
Agent Hub admin tool list. The left menu lists dashboard, tools 16, skills 22, import installed items, MCP servers, agent connections, shared notes, call log, session log and devices; the middle lists shared tools such as hub_status and hub_search_skills with their descriptions.
16 shared tools · only paths and device names covered
Agent Hub skills screen. Skill cards such as agent-dispatch and competitor-profiling sit in two rows with a description, version and the agents that use them.
22 shared skills
Agent Hub agent connections screen with cards for Codex, Claude Code and Cursor and an MCP settings export panel. Config file and program paths are covered with black bars.
Connection settings per agent
03

AND

Autonomous Neuro Drive2026.08 –In use

LLM wiki2026.08 –401 Markdown pagesNavigated by index.mdLimit: about 100 sources2026.10 switchSQLite as the single sourceAI conversationsAgent Hub notesWork logsGit historyraw_eventsImmutable sourceExtract, verifyverified_claimsLinked to evidencesearch_indexFTS5Morphemes (Kiwi)Vectors (sqlite-vec)MCP serversearch, ask, timelineWikiDerived view

AI chats, work logs, notes and Git history go into one database, and facts with checked evidence are pulled out separately. From any agent I can ask “what did I do back then?” and get an answer with its sources.

What was my role in the Bearable project? And what did I say in meetings?

Example AND answer

From August to September 2026, you were team lead of a five-person team on the MyFolio PoC with the edtech company Bearable. You built and deployed the university information and high school course roadmap screens, ran the repository, and coordinated requirements with the company.

  • Weekly meetings: you took the company's requests, laid out which screens the next sprint would build, and shared that with the team.
  • September 23 presentation: you presented the course-planning agent design and the measured call costs, and said the company's feedback would go into the next version.

Sources: 3 verified facts · 2 chat logs · 1 slide deck

Search
Under 1 sMorpheme and vector search, 10 results
Writing the answer
About 5–10 sLocal model on the desktop GPU
Cloud tokens
0The whole answer stays on my own machine
Asked from a coding agent
About 2–4k tokensFor reading and summarizing the results, 10–20 s

Rough figures. When no evidence is found, it says so instead of answering.

Questions with a correct answer in the top 8, out of 30 (2026.11.09)

  1. Lexical search24/30
  2. Lexical + vectors26/30
  3. With LLM reranking30/30
  • 18,579Search index rows
  • 18Migrations
  • 21ADRs
  • 312Tests

Lexical search plus semantic search, then a final reranking, put the right answer in the top 8 for all 30 questions.

04

Grok bot and agent automation

Personal automation2026.08 –In use

Claude Code, Codex, CursorAdd to-dosAgentHubUpdate and spare-token noticesClaude Code, CodexDelegate jobs (MCP)WebhookGrok botCloud routineRuns after approvalResults returnedTo-do listDeduplicatedDeadline reminders1 day and 1 hour beforeMorning briefingSchedule, mail, newsUnchecked itemsCarried to the next report

When I say “add this to my to-dos” in the middle of a chat, the Grok bot takes it, files it, reminds me before the deadline and briefs me on the day every morning.

  1. Ask in plain words

    While working with a coding agent, I say “Toss documents due Friday 6 pm, add it to my to-dos.” The agent turns that into one to-do with a title, deadline, urgency and related link.

  2. Hand it to the bot

    A small assistant app on the Mac sends it to the Grok bot over a webhook within 10 seconds. Anything that looks like a password or card number is hidden first, and failed sends are retried at growing intervals.

  3. File it

    The Grok bot adds it to my to-do list and converts the deadline to Korean time. The same to-do arriving again updates the old one. Instructions hidden in the text are written down, never followed.

    To-do added: Submit Toss documents (due Fri 11/13 18:00)

  4. Remind on time

    It reminds me a day and an hour before the deadline, and right away for urgent items. The morning briefing brings the day's schedule, mail and news, and anything unchecked moves to the next report.

A public copy of how the Grok bot connects to the Mac, the laptop and the cloud. Text is covered with black bars. Open full size
MacBook AirMain workSamsung laptopAlways-on serverDesktopGPUJob dispatch (SSH, Tailscale)Results returnedWake-on-LANGPU compute (SSH)Headless runs: cursor-agent, grok, codex, claudeScheduled jobs (one hour):00:10:20:30:40:50:60CollectProcessSpare-token jobs (every 10 min)

Long jobs go to the always-on laptop or the GPU desktop, and the Mac only collects the results. If the desktop is off, it is woken remotely.

05

Other projects

4 team projects7 personal tools

  • Myfolio high school credit roadmap screen. Elective course cards are arranged in columns by grade and semester.

    Bearable PoC

    2026.08 – 09Team of 5Team lead

    Built and deployed a high school course-planning screen and a recommendation agent for an edtech company's service.

  • Publium search and answer screen. PubMed search filters on the left, a list of paper chat rooms in the middle, and an AI answer based on the selected papers on the right.

    Publium

    2026.07Team of 2, 8 daysBackend, search agent

    A research search service that finds PubMed papers and answers with the paper IDs as evidence.

  • Beorimi start screen. A photo button sits under the line “How should I throw this away?”Detection result. Numbered boxes mark four items in the photo, chairs and a chest of drawers.Report preparation screen. Under the title “I’ll get the report for 3 items ready” are the expected fee per item, a note field for disposal details, and a button to the official report page.

    Beorimi

    2026.08Team of 3, 28 hoursTeam lead, backend

    From one photo it finds several items to throw away and prepares each item's fee and the disposal report.

  • Diary writing screen. The title, weather and entry are typed in a handwriting style.Converted result. A crayon drawing of a waterfall, the entry written on manuscript paper, and a teacher’s stamp.Byeolil Observatory main screen. An orange orb floats on a dark space background, with a line inviting you to take out the fortune inside it.

    Two Apps in Toss mini apps

    2026.07 – 08Officially launchedPlanning lead, team lead

    Released two mini apps, My Summer Vacation Diary and Byeolil, on Toss Apps in Toss.

  • Galpi

    A navigation tool that helps coding agents find the right place while reading less code. Measured with A/B tests.

  • Poom

    Hands small jobs to local models and checks the results before taking them.

  • AQR

    Queues AI jobs by priority and order, runs several side by side and returns the results as PRs.

  • Nabiseo

    A personal assistant that operates my phone. It always stops at the payment step. (private)

  • FAPA

    Collects card statements automatically and sorts the spending. (private)

  • Momento

    A tutoring schedule app shared by tutor, student and guardian. Built and run on my own.

  • cogito

    A personal agent for testing how to split work between an agent and plain code. (private)

Contact

jae1450@naver.com

Job offers and teaching inquiries are both welcome.

© 2026 Jason

  

Background: matrix rain and a hologram face (my own dynamic-css-motion skill). Face model: MediaPipe face mesh (Apache 2.0). Expression tracking: MediaPipe Face Landmarker (Apache 2.0). License: MediaPipe