AI agent portfolio

Creative, original
infographic videos

Chart races, 3D terrain flights, a Blender 3D world map, pitch heatmaps. For each topic I design how the data should be seen, starting from the format itself, to make videos you will not find on other channels. Data comes from more than 20 kinds of sources, from sports records and economic indicators to opinion polls, streaming charts and terrain elevation, and the repeatable work (collection, aggregation, rendering) runs on a pipeline built with AI agents (Claude Code). The same templates also power English-language channels as the work expands overseas. One frame from one video per type is shown here. Click a frame to open that video on YouTube.

  • 28infographic video types
  • 10+visualization templates I designed
  • 20+data sources
  • 56featured videos by type
Pipeline

How one video gets made

  1. SRCSourceAPIs · sites · public stats
  2. GETCollectPython · Playwright
  3. DBLoadMSSQL · DuckDB
  4. SQLAggregateTop-N by period · matrices
  5. MP4RenderRemotion · TTS · thumbnail
  6. PUBPublishmetadata · channel routing
  7. KPIMeasureview stats → next topic

Where the data comes from

  • KBO box scores and play-by-play
  • MLB · NPB
  • EPL · K League · internationals
  • KBL · NBA
  • KLPGA prize money
  • Football events and heatmap coordinates
  • Player market values and transfers
  • FIFA rankings
  • TV ratings
  • Streaming Top 10
  • Box office
  • Real estate transactions
  • Company financials and market caps
  • National economic indicators
  • Opinion polls
  • News big data
  • YouTube channel stats
  • Satellite imagery and terrain elevation

Who does what

Person · planning and directing

  • Designs a new format for how the data should be seen
  • Directs the scenes, colors and motion
  • Chooses what to make (topic board)
  • Reviews numbers and wording
  • Turns each agent mistake into a rule

Agent · the repeatable work

  • Collects and loads data daily from 20+ sources
  • Aggregates with queries and writes the script
  • Renders video, voice, thumbnail and metadata
  • Gathers view stats and proposes the next topic
By type

28 types, 56 videos

Each type started with a new format. Once built, a template is reused with its query skeleton, so a new topic only changes the data.

01

All-time bar chart races

Decades of cumulative records replayed as rankings change, with side panels such as "compared at the same age".

02

Daily rank trajectory races

Every day of the season joined into rank lines beside the bar race, showing who climbed and when.

03

Treemap races

A bar race and a treemap grow together, showing rank and share on one screen.

04

YouTube channel races

Monthly channel views turned into a race, with the hit video behind each month's leader.

05

Head-to-head matrices

Team matchups re-aggregated by date into a matrix, updated daily during the season.

06

Nemesis analysis

Every hitter-versus-pitcher record matched up to find who dominated whom.

07

Long-form match reviews

After the final whistle: box score, batted-ball and shot locations, scoring by inning or quarter, published the same day. Baseball, football and basketball share one skeleton.

08

Review and preview shorts

The same match data, directed again as a short. Previews lead with the starting matchup and head-to-head numbers.

09

Weekly reports

A whole season sliced into weeks: each week's Top 5 and standout player on one event timeline.

10

Football heatmaps

Player location data drawn on real pitch coordinates.

11

Basketball shot charts

Shot locations and makes plotted on the court, one three-point map per player.

12

Blender 3D world maps

Blender driven from Python (bpy): bars rise country by country while the camera flies over the map, all generated from data. One builder produced shorts for all 12 World Cup groups.

13

Blender 3D bar chart races

Item count, bar heights and camera paths are computed from the data and rendered as resumable PNG sequences.

14

Blender 3D ballpark

The camera walks through a 3D ballpark of pillars, one per year, meeting the star of each era.

15

Blender stat duels

Eighteen stats from one match, the ranking rebuilt for each.

16

3D terrain flights · fighter jet

Satellite imagery and elevation data are downloaded along the route and baked into 3D terrain (about 9 GB for Jeju), and an F-15K flies low along a hand-picked route. The aircraft was repainted and its rig baked in Blender.

17

3D helicopter tours

A helicopter follows real roads or a list of mountains, with speed, altitude, place names and a minimap computed for every frame.

18

3D map rankings

Mountains raised in 3D on a district map and visited one by one in order of height.

19

3D pitch and home run replays

Exit velocity, launch angle and pitch tracking data rebuild the moment in 3D, down to whether the homer would clear other ballparks.

20

Season ranking cards

One ranking table as a 7-second short: a single query feeds a card template that sets title, highlights and colors.

21

All-time cards (Korea, US and Japan combined)

Career rankings re-computed with overseas league records added.

22

Dream team lineup cards

The best player at each position, picked by high school, laid out as a lineup.

23

Economy and country indicator cards

The agent scans news and public statistics daily; a person picks a topic, the numbers are verified, the card is built.

24

Streaming and YouTube ranking cards

Streaming charts and YouTube channel stats turned into ranking cards.

25

Survey and poll cards

Polls and surveys re-cut by age, region or country.

26

VS comparison cards

Two players or two seasons, side by side in one table.

27

Stat quizzes

A quiz series where one line of stats identifies a player, drawn from the same records database.

28

Expanding to overseas channels

The same data and templates with a different language dictionary now run English-language channels (Football Report and Hoops Report). Match reviews ship in Korean and English on the same day.

How it is made

This page was built the same way

Records are the source of truth

Every video has a manifest with data sources, composition, render version and upload URL. The list on this page was generated from those records.

Frames, not thumbnails

Each image is a single frame cut from the video file with ffmpeg: the actual screen, not a thumbnail.

Run by rules

Each project carries an operating-rules document; hooks block paid calls and visible console windows before they run. Every incident becomes a rule.

Tools

  • Claude Code
  • Python
  • Playwright
  • MSSQL
  • DuckDB
  • Remotion
  • Blender
  • d3
  • Three.js
  • ffmpeg
  • TTS
  • gpt-image
  • YouTube Data API
  • Task Scheduler
  • Cloudflare Pages

Need a pipeline like this?

From data collection to automated videos and reports, designed as a system one person can run.

Get in touch