Jisang Han

01 I started with hardware

Before SLIT, I built a portable braille reader. It read printed text with OCR and pushed the letters out through solenoids. It took us to the finals of a youth startup competition.

What stayed with me was simple. A thing becomes real when someone can pick it up and use it without you standing next to them.

02 Automation is a habit

When I see a task repeating, I start writing the thing that will do it next time. My Threads account grew that way, to about seventy thousand followers. Most of the work behind it runs on systems I built: they collect what performed, find the pattern, draft the next post and publish it.

I also keep a startup diary on Instagram, @physicat_official, read by about seven thousand people. kakao-headless came from the same itch: agents should be able to use the tools people actually live in.

03 An engineer at the desk

At SLIT I own what we build and how it runs. Most of that is Recova, our agent for debt recovery. It reads court rulings, property registries and credit records, then suggests how recoverable each account is and what to do next. It is in pilots now.

Prompts alone did not get us there. Collections runs on the judgment of experienced people, so I sat next to them and watched what they checked, and in what order. Inheritance, competing claims, statutes of limitation: the exceptions pile up, and one standard has to hold across a whole ledger. We rebuilt the documents and ran quality loops until it did.

The agent now explains why an account deserves another look and whether it is worth the cost. People approve and negotiate. The agent can prepare. People decide. My job is to make that line clear in the code.

04 Beyond Recova

The pattern travels. For a rent-a-car company, the same structure handles traffic fines, from looking up each case to processing it and checking the result. For a law firm, a system that drafts answers to consultation requests shipped in two weeks.

I also run marketing automation for a company whose apps have passed thirty million downloads. A bot gathers each channel’s results every day, finds what worked, makes the next copy, card news and video, publishes it, and lets the results shape the next round. Meta ads run through the same loop.

05 Fixing the tools I use

I maintain cli-jaw, an open-source agent runtime, and run projects of my own: kakao-headless, neuralvault for self-hosted AI memory, k-spy for Korea’s public CCTV feeds, and a multi-agent suite that splits planning, design, engineering and the backlog.

Upstream, my fixes have been merged into axios, Redux, Pinia, XState, aiohttp, promptfoo, fastmcp and oh-my-openagent, among others. Almost every one started as a bug I hit while using the tool.

06 All in

I study Electrical and Electronic Engineering at Sungkyunkwan University, and I am on leave to build SLIT full time. Two grand prizes with ANYON, at the AI+X Global Startup Competition and Gyeonggi Deep Tech FLEX 2025, were the push to stop splitting my days.

What I do has not changed: find the work that repeats, build the agent that runs it, and stay until the numbers move.