Brian Shin

Ten years of my golf, analyzed

Tue, 06 Oct 2026 00:00:00 +0000

Since 2016 I’ve logged almost every round of golf I’ve played, first in Swing by Swing (now SwingU) and then in 18Birdies, and until this week I’d never looked back at any of it. That’s 194 rounds across Colorado, California and Korea. I also had a summer of range sessions from 2024 sitting in my Toptracer account.

So I exported all of it and spent a long session with Claude Code turning it into a dashboard. It’s live at tools.brianshin.me/golf.

One of my favorite pieces is this chart: my handicap over time, built like a stock chart, so I can scrub through any date or drag the window underneath. My index went from 19.2 (when I first had enough rounds for one, in 2020) to 10.5 today. The spike in early 2021 is me coming back from a long layoff, and the drop after it is the 64 rounds I played that year.

Building it

The first job was getting the data out, and that turned out to be the hard part. I tracked my early rounds in Swing by Swing, which later became SwingU, then switched to 18Birdies at the end of 2020. I honestly don’t remember why. I think SwingU was draining my phone’s battery. None of these apps make it easy to take your history with you. 18Birdies’ came out as an archive of my whole account, and SwingU’s as a raw hole-by-hole file. Toptracer has no export at all; after some digging I found SwingSync, a third-party service that connects to your Toptracer account and pulls every session out as a spreadsheet.

After that, I mostly described what I wanted and judged what came back. Claude did the building.

The data was messier than I expected. 18Birdies doesn’t export the par for each hole, so Claude tracked down the scorecards for the 35 courses I’ve played and matched them to my rounds. It also caught a round I’d logged in both apps.

One disappointment: 18Birdies’ export only includes the basics. It has my score on every hole, but just round totals for fairways, greens in regulation and putts. The detail I logged hole by hole (putts, chips, sand shots and penalties) didn’t come out at all, which is why the dashboard says so little about my short game. The app clearly has that data, so I’m working on getting it out another way, and I’ll add it when I do.

I’m a big fan of Data Golf, so I asked Claude to look through their site and borrow what made sense for an amateur. The result is a page that estimates what I’ll actually shoot next time out (a few strokes worse than my handicap suggests) and forecasts my next round at each of my regular courses.

The part I enjoyed most was the maps. Most of the Bay Area courses I play have their holes traced in OpenStreetMap, so every round now has a replay: step through it with the arrow keys and each hole is framed on satellite imagery, with my score, what I’d be expected to make, and every other time I’ve played it.

What my rounds said

My favorite page is the game plan. Every finding comes with the evidence, something concrete to try, and a rough estimate of how many strokes it’s worth.

Stop making doubles. 32% of my strokes over par come from going beyond bogey, almost 6 a round. My bad rounds don’t have more bogeys than my good ones. They have more doubles: 6.9 a round on my worst days against 3.7 otherwise. The fastest way down for me isn’t better shots, it’s fewer disasters.
Take one more club. 46% of my missed greens come up short, and only 13% long.
Practice approaches, not driver. For me, hitting more greens is worth far more than hitting more fairways.

I also compared my averages with Shot Scope’s numbers for a typical 10 handicap. Over my last 20 rounds I hit greens about as often as they do (36% against 35%), but fewer fairways (40% against 50%), and par 4s are where I give up the most: about 3 strokes a round.

It also found that I play better before 10am, and that hole 16 at Half Moon Bay’s Old Course has my number. None of it is shocking golf advice. But seeing it in my own numbers, ranked by what it’s worth, is a lot more motivating than reading it in a magazine.

What the range said

In August and September 2024 I hit 1,994 balls over 29 sessions at a Toptracer range, and SwingSync pulled every shot out with its carry, height, launch, spin and curve. Range balls fly a little differently from the ones I play on the course, so the exact yardages are probably a bit off, but the patterns are clear.

My typical 7 iron carries 157

On a good strike, my 7 iron carries 157 yards. Half of them land between 149 and 164, only one in ten goes past 170, and my longest was 186.

Four clubs, one distance

My 3 hybrid, 4 iron, 5 iron and 6 iron carry 178, 175, 169 and 165. Four clubs inside about 12 yards, where you’d want 10 to 15 between each. I’ve been carrying one club with four numbers on it. From the 6 iron down the gaps are textbook, and the wedges are the best part of my bag.

Here’s how high the ball gets, club by club, next to tour averages:

From the 6 iron down, my shots get about as high as an LPGA player’s. Above that they fall off a cliff. My 4 iron peaks at 15 yards; a tour player’s peaks at 28. It also comes down at a much shallower angle: 29° against about 48° for a tour 4 iron.

More ball speed than an LPGA player, less carry

My driver ball speed is higher than the LPGA average (147 mph against 140), but I carry it 211 yards to their 218, because they hit it higher. Launching it like they do would be worth roughly 7 yards without swinging any harder.

The big miss is a hook

55% of my full swings curve left. With irons that’s mostly a draw. With the driver it’s a hook: 20% of my good drives finish more than 30 yards left, against 9% more than 30 right.

Interestingly, my logged fairway misses on the course split about evenly between left and right.

The first 40 balls are the worst

In the first 40% of a session, 10% of my swings are mishits. After that, under 4%. On the course you don’t get 40 balls, so the first tee gets the cold swing.

The data also showed how I practice: about 90 balls a session, hit in runs of 8 with the same club. The standard advice is to switch clubs and targets every ball instead. Yes, I know. Seeing it in my own numbers makes it harder to ignore.

Straighter, but less clean

Over the summer I got straighter with most clubs (the 6 iron’s spread went from 19.5 yards to 14.3), but my mishits went up at the same time (the 7 iron from 3% to 12%).

What I’m changing

Club by my real carries. Put them on a card in the bag, and between clubs, take the longer one.
Fix the top of the bag. Four clubs for one distance is two too many. Swap some of them for clubs that launch higher, and have a fitter check the gaps.
Launch the driver higher. Tee it higher, ball forward, sweep up through it.
Play for bogey on the hardest holes, where most of my doubles happen.
Warm up before rounds that matter.
Mix up range sessions. Instead of hitting the same club over and over, switch clubs and targets every few balls, the way a round actually goes.

What’s next

First, the hole-by-hole data that 18Birdies didn’t export: putts, chips, sand shots and penalties on every hole. With that in, some of these findings may change, and the dashboard can finally say something useful about my short game and putting.

After that, I’d like hole maps for Spring Valley, my most-played course, and for my courses in Korea.

Separately, I’ve been working with Claude on editing my old range session footage and connecting it to my Toptracer sessions, so each clip shows where the shot went and how far. Subscribe to my YouTube channel if you want to catch those.

I do know what any golf coach would say. The old rule is to practice your short game twice as much as your long game, and here I am with 1,994 range balls and a whole dashboard about them. But this was about looking at what I actually do, which is hit buckets of balls at the range and walk onto the first tee without warming up. The short game is next. Probably.

If you track your rounds or range sessions and want something like this for yourself, let me know.

kakaocli: letting AI agents use KakaoTalk

Mon, 05 Oct 2026 00:00:00 +0000

I made kakaocli, a command-line tool for KakaoTalk on macOS, so that AI agents like Claude Code can read, search, and send my KakaoTalk messages. I’ve also added it to my tools page.

In Korea, pretty much everything happens on KakaoTalk, but Kakao’s official APIs can’t read your chat history or send a normal message. So I couldn’t ask an agent to catch me up on a busy group chat or draft a reply, which is exactly the kind of thing I want agents doing for me.

kakaocli gets around this without touching Kakao’s servers. It reads the encrypted database that KakaoTalk for Mac already keeps on your computer (read-only), and it sends messages by driving the real app through macOS’s accessibility features. Everything has JSON output, so an agent can just call it. Getting started is one line: brew install silver-flight-group/tap/kakaocli.

I built it back in February and open-sourced it under MIT, and it’s been fun to see it pick up over 150 stars and 60 forks on GitHub. It’s Mac-only and still text-only, so if you try it, start with your own 나와의 채팅 self-chat before letting an agent send anything to real people.

Logistics Tech Review is back

Mon, 23 Mar 2026 00:00:00 +0000

I’ve relaunched Logistics Tech Review, my newsletter covering technology, automation, and market shifts across logistics and supply chain. I started it back in 2020 while I was working on autonomous vehicles and robotics, took a break, and have been publishing weekly again since earlier this year.

The newsletter covers the intersection of technology and the physical supply chain — things like warehouse robotics, autonomous trucking, port automation, geopolitical shipping disruptions, and how software platforms are reshaping freight. If you’re in logistics, investing in the space, or just interested in how physical industries are being transformed by tech, I think you’ll find it useful.

Some recent issues:

No Bypass, No Drivers, No Plan B — on Hormuz disruptions, diesel, and CDL enforcement
The $12 Billion Bet on Logistics Consolidation — Thoma Bravo’s shipping acquisition and Supreme Court broker liability
Physical AI’s Billion-Dollar Bet — humanoid robots in manufacturing and warehouse automation

You can subscribe at logistics.substack.com. I’m also on Twitter at @LogisticsReview for the newsletter and @brianshin22 personally.

Adding tools.brianshin.me

Fri, 25 Oct 2024 00:00:00 +0000

I created a projects page at tools.brianshin.me to share small projects I create. This is inspired by Simon Willison’s blog and his tools page. I’ve started creating a few tools using Claude’s artifacts tool, which is amazing. Going forward, I hope I can use this page to host interesting tools and ideas that aren’t limited to just a single-page HTML/JS page.

I’ve made a bunch of things now with Claude, but so far the ones I’m putting on the tools page are:

Youtube Translate landing page A sample landing page I’m working on.
Annual Return Calculator A quick page to help calculate an investor’s compound annual return, given starting, ending, and withdrawal/deposit amounts. I made this because I find myself wondering this more often than I’d like to admit and wanted a quick, no-frills place I could go to calculate it.

This new workflow of using Claude’s artifacts tool and the model behind it (Claude Sonnet 3.5 new today) feels like a game-changer. I’ve been using it heavily for my main work in coding now, and I can’t really go back. It’s one of those things where, once you know it exists, you can’t go back to a world where that knowledge doesn’t exist for you. Kind of like having experienced seeing color for the first time I guess. It’s also great that I can create new tools from my iPad or phone, and see the results. It’s like magic - I’ve been going through my running ideas doc and literally having Claude spin up my ideas for me in less than five minutes.

I want to try some of the more advanced tools where you can host React apps that Claude builds: claude-artifact-runner. Of course, I hope it isn’t long where I can ask AI to build full web apps with full-stack functionality and deploy them as trivially as I have with these html pages.

I analogize what’s happening today with these kinds of tools as similar to when the printing press was invented, or when computers and specifically word-processing/typing came out. It is like a step-change in enabling everyone to create or write, and consequently this has second-order effects on consuming the writings and increasing overall productivity.

Of course, I think real business value will always come from focusing, iterating and refining a solution to a problem, and solving real people’s needs. In that sense, the ability to quickly prototype is very valuable but not as valuable as building a product or products as a business over a long period of time. I guess it would be like the proliferation of printing and typing/word processing enabling the best writers, publications, important works to be distributed and enjoyed even more, and not necessarily killing any advantages they had simply because they could write well already.

This tools page was relatively painless to set up. Even though my main website is now on Gitlab Pages, I set up this tools subdomain on Github pages, and it was easy to set up the custom domain edits on Cloudflare DNS.

CES 2021

Sun, 22 Aug 2021 00:00:00 +0000

CES 2021 was virtual due to COVID-19, but our team's work, LGSVL Simulator, was featured in the LGE CTO's LG Future Talk. This was very cool, as it was one of only a few projects highlighted as part of LG Electronics' broader business and technology initiatives at CES. You can see LG's CTO I.P. Park talk about it below, as well as the video demos our team created.

CES 2020

Sun, 22 Aug 2021 00:00:00 +0000

At CES 2020, we showed off LGSVL Simulator at the public Automotive Grade Linux exhibit once again. We gave demos of advances in LGSVL Simulator, including new capabilities like being able to procedurally generate environments using VectorZero's RoadRunner tool, then importing to our simulator.

Additionally, at least three other companies at CES 2020 were demonstrating the LGSVL Simulator technology in their respective booths, a sign of growing adoption!

At the award-winning AWS Automotive exhibit, Unity showed off their product, Unity Simulation, demonstrating LGSVL Simulator being scaled and run in the cloud. You can read more about it here.

Apex.AI also used LGSVL Simulator to demonstrate their product, Apex.AI and Apex.OS. You can watch their CES 2020 video below.

DXC also demonstrated their offerings, and demonstrated LGSVL Simulator technology in their booth. You can read more about Robotic Drive here.

ROSCon and IROS 2019

Sun, 22 Aug 2021 00:00:00 +0000

ROSCon and the IEEE/RSJ International Conference on Intelligent RObots and Systems (IROS - one of the two biggest robotics conferences every year) were held in Macau in 2019.

We went to give a talk at ROSCon, as well as demonstrate LGSVL Simulator and webOS Robotics at exhibitor booths in both conferences. You can see my presentation at ROSCon on my presentations page.

You can see our prototype robot at our booth in IROS. Though the hardware looks very prototype-ey, the software is pretty impressive. It is a multi-display touchscreen user interface, algorithms to detect obstacles and faces, ROS 2 and teleoperation control, and voice recognition, all running on one single Raspberry Pi 4.

Today, a lot of autonomous mobile robots deployed commercially have very powerful workstations on them running Ubuntu - we demonstrate the advantages of using a lightweight embedded operating system on much cheaper hardware to meet all the typical use cases of a mobile robot.

We also demonstrated LGSVL Simulator, for developers working on autonmous vehicles or robotics needing simulation tools.

Unity Unite 2019

Sat, 21 Aug 2021 00:00:00 +0000

At Unity's annual Unite conference in 2019, Unity Simulations was debuted. Unity Simulations is made for running cloud simulations for applications based on Unity, focused on industrial applications such as automotive, transportation, and robotics.

You can see part of the keynote featuring LGSVL Simulator in the video below.

SVL Simulator

Sat, 21 Aug 2021 00:00:00 +0000

SVL Simulator (formerly LGSVL Simulator) is an autonomous vehicle and robotics simulator. It is for engineers and developers building autonomous vehicle (AV) software or robotics algorithms, which they can use to test and verify their systems. It provides a real time 3D virtual environment to provide the dynamic environment, the “world”, that the AV or robotics stack perceives through its sensors. It is based on the Unity game engine.

It’s a product that is free to use, and open source on Github (with LG’s own license). This is what I have spent most of my time building at LG for the last three years or so, and this page shows off more about what it is and the public impact it has had in the autonomous vehicle and robotics spaces.

Who Uses It

SVL Simulator is used by:

Engineers at companies building autonomous vehicle and robotics software and systems
Engineers at companies developing parts of or technologies for autonomous vehicle systems: sensors, hardware-in-the-loop, engineering services
Researchers and students investigating autonomous vehicles, robotics, machine learning, and many other applications such as security, future mobility, and new programming languages

Some numbers

~3000 monthly active users around the world
Downloaded over 50,000 times since our launch in November 2018
Usually at least 20 simulations being run at any time globally!

Research

Below are links to some highlighted papers referencing the SVL Simulator by LG.

LGSVL Simulator: A High Fidelity Simulator for Autonomous Driving - Our own paper published in IEEE International Conference on Intelligent Transportation Systems 2020 giving an overview.
Formal Scenario-Based Testing of Autonomous Vehicles: From Simulation to the Real World - A scenario-based testing framework and scenario description language, using SVL Simulator.
Safety Assessment and Simulation of Autonomous Vehicle in Urban Environments - a safety evaluation of a real autonomous vehicle shuttle using our simulator.

Public projects using SVL Simulator

The 2021 IEEE Autonomous Driving AI Test Challenge - a challenge for building testing frameworks for autonomous vehicles using SVL Simulator as the competition platform to be evaluated against.
F1Tenth - A competition for small-scale autonomous racing vehicles, using SVL Simulator for simulation
Autoware Foundation Autonomous Valet Parking 2020 - A hackathon to demonstrate an autonomous valet parking operational design domain (ODD) use case, first in simulation then in the real world.
Special Topics in Autonomous Driving: A graduate engineering class at San Jose State University in Spring 2020 that adopted SVL Simulator as part of its curriculum.

Presentations

Sat, 21 Aug 2021 00:00:00 +0000

ROSCon 2019, Oct. 2019. “ROS on OpenEmbedded: Simpler robotics development for embedded Linux”

My colleagues and I gave a talk at ROSCon 2019 in Macau presenting the advantages of using embedded Linux for robotics development. We also demonstrated a real robot at the conference showcasing the use cases. Slides.

AI Robotics KR, Oct. 2019. “ROS2 on webOS”

Presentation to 150 audience members at AI Robotics meetup in Seoul about LG’s webOS Robotics intiative and open source robotics work. Slides.

Silicon Valley Autoware Meetup, Sep. 2019. “LGSVL Simulator: Advanced Sensor Rendering for Autonomous Vehicle Simulation”

A presentation given by a colleague and myself on LGSVL Simulator and its sensor simulation for a Silicon Valley Autoware Meetup attended by 50 people, hosted by Apex.AI in Palo Alto. You can see parts of the live demo and Q&A session in the video below.

ROSCon 2018, Oct. 2018. “ROS2 on webOS: Web-app enabled robots”

I gave a lightning talk at ROSCon 2018 about the Beanbird bot prototype demonstrating LG’s embedded operating system webOS running ROS2. This was a precursor to our full presentation a year later at ROSCon 2019.

Brian Shin

Ten years of my golf, analyzed

Building it

What my rounds said

What the range said

My typical 7 iron carries 157

Four clubs, one distance

More ball speed than an LPGA player, less carry

The big miss is a hook

The first 40 balls are the worst

Straighter, but less clean

What I’m changing

What’s next

kakaocli: letting AI agents use KakaoTalk

Logistics Tech Review is back

Adding tools.brianshin.me

CES 2021

CES 2020

ROSCon and IROS 2019

Unity Unite 2019

SVL Simulator

Who Uses It

Some numbers

Research

Public projects using SVL Simulator

Links

Presentations

ROSCon 2019, Oct. 2019. “ROS on OpenEmbedded: Simpler robotics development for embedded Linux”

AI Robotics KR, Oct. 2019. “ROS2 on webOS”

Silicon Valley Autoware Meetup, Sep. 2019. “LGSVL Simulator: Advanced Sensor Rendering for Autonomous Vehicle Simulation”

ROSCon 2018, Oct. 2018. “ROS2 on webOS: Web-app enabled robots”