Rack Labs, Seoul

An AI-native startup
that ships hardware.

Rack Labs makes RackTracker, a physical product for gyms: an edge device with three cameras and an NVIDIA Jetson that mounts on the power rack, and the software that turns every set into biomechanical feedback. We build all of it with Claude Code, and we open-source the tools we make along the way.

RackTracker Open source

RackTracker

Our main business. The rack that measures the lift.

RackTracker is our main business, and it is a physical product: a hardware edge device and the software that runs on it, supplied and installed as one unit per power rack. The device is three global-shutter cameras and an NVIDIA Jetson fixed to the rack.

The cameras fire on one electrical trigger. The joints found in each view are intersected from three directions, and when one view is occluded the other two carry it, which gives cm-level 3D coordinates.

Every set is logged as center-of-mass shift, left-right asymmetry, trunk lean and squat depth, and handed back to the member and the trainer. Thresholds are set per person from body proportions.

Because the setup never changes, a new session compares with the last one on the same footing. Processing finishes on the computer beside the rack, on a closed network, so member video never leaves the gym.

What we supply: Edge device + softwareone unit per rack, installed by us
Shutter skew across three cameras: 0.12 msmeasured with a logic analyzer
Joints recovered with three cameras: 70%CMU Panoptic. Standard triangulation: 40%
Joint position error: ~3 cmthe same as with two cameras
Where processing runs: NVIDIA Jetson beside the rackclosed network, no video leaves the gym

Read the business plan

Analysis view of a squat at the bottom position: joint angles, per-joint load and the center-of-mass line drawn over the lifter — Figure 1. Joint load and center-of-mass analysis.

NVIDIA Jetson compute unit on the bench with the trigger board and camera modules wired to it — Figure 2. Jetson edge rig with the trigger board.

NowWorking prototype: three-camera capture, calibration, 3D joint reconstruction and automatic set logging
Q4 2026Validation against marker-based motion capture
H1 2027Gym pilot
H2 2027Production unit

Prototype

Photographed 17 September 2026 on the PVC rack mock-up

PVC power-rack mock-up in a lab room with three cameras clamped to the uprights and a tablet on the bench — PVC rack mock-up, three cameras mounted

ChArUco calibration board leaning on the rig beside a laptop running the kiosk — Calibration board on the rig

Live camera view in the kiosk with a green grid while a person holds the calibration board — Kiosk camera view during calibration

Hand-built trigger board on a breadboard, held over the Jetson with its jumper wires — Trigger board, one pulse to three cameras

Jetson module with its heatsink on the carrier board, case opened on the bench — Jetson module on the carrier board

Cameras: 3 AR0234 global-shutter sensors, 120-degree low-distortion lens
Compute: NVIDIA Jetson, TensorRT inference
Trigger: Hand-built board, one pulse to all three cameras
Mock-up: PVC power rack, cameras on magic arms

Open source

Tools we build along the way, published for other teams

Two projects come out of building RackTracker with Claude Code. pipeplot is live now, and its code will be released as open source. agent-workflow-kit is already on GitHub under the MIT license.

pipeplot: Open-source release plannedlive at pipeplot.pages.dev
agent-workflow-kit: MIT, v0.2.13github.com/lhk0721/agent-workflow-kit

pipeplot, open-source release planned

Neural networks,
built layer by layer
in walkable space.

One page per model. Model pages stand real trained weights and activations up in 3D space, and concept pages stack scenes in the order one person came to understand the idea.

Four models so far, AlexNet, VGG-16, GoogLeNet and ResNet-34, plus concept pages. It is growing into a place to learn ML and DL, and a community around it. The code will be released as open source.

Open pipeplot See the models

agent-workflow-kit

Open source under MIT. A workflow, a Markdown document store and backstops for repositories where Claude Code does the work

Agent Workflow Kit gives a repository three things: a workflow that turns a request into an issue, a branch and a document, and runs many of them at once on git; a document store kept in Markdown and engineered so the agent's context stays small and true; and backstops that fire when a step was skipped.

This repo is the single source: each target repo gets git-tracked copies through install.mjs, so a fix made here reaches every repo on its next update instead of drifting per copy.

Version: 0.2.13updated with the release notes
Skills: 9issue-start, post-pr-cleanup, ui-evidence, ko-writing and more
License: MITNode 20 or newer

View on GitHub

Flow of one work unit: issue, branch and worktree, management document and pointer, then edits, review gate, commit, and PR with cleanup — Figure 3. One work unit, from issue to cleanup.