Junyi is building.
I build products, train models, and run the machines underneath. This is where I write about what I made, the numbers I measured, what failed, and the parts I still think are genuinely cool.
Things I have built
Selected work
Four things I have spent enough time with to know where the clean description stops and the real story starts.
- 01Product · travel information
Piaopiao Travel
A travel planner that covers visa rules for 199 countries and regions, wired to real flights, hotels, and activities.
I built it from the data up. Every visa entry traces back to an official source, and official sources are reviewed through a fixed weekly process. Flight and hotel results come from live inventory, not cached guesses.
One country's official site went down and stayed down for ten days. The monitoring knew before anyone asked.
- 02Positioning · information design
Water technology
An R&D company in water treatment. I own its outward-facing layer: how its work is explained, organized, and shown.
The same verified facts have to read correctly to an engineer, a regulator, and a customer. Most of the job is deciding what a number needs next to it before it can be trusted.
The venture isn't mine to name here. The craft is.
- 03AI · measurement
Model training and evaluation
I fine-tuned an 8-billion-parameter model on a single RTX 3090 until it matched a commercial model on my own benchmark: 0.923, exactly.
Then I grew the training set about six-fold, to 28,810 examples. The score did not move by a single point. With thirteen binary cases, the benchmark could only resolve differences of about 7.7 percentage points. My own weights cleared the release gate and never served a user; the RTX 3090 belonged to a training experiment that is now retired.
The training machine gave up before the questions did.
- 04Operations · reliability
Independent infrastructure
The things above run on infrastructure I operate myself, with independent monitoring, alerting, and rehearsed recovery.
The rule is simple: nothing gets to report its own health. A separate layer decides what is actually up, recovery paths are exercised, and failures become operating notes instead of private memory.
In all these years, not one machine has ever announced its own failure. The monitoring exists because of that.
Notes from the work
Writing
Longer pieces about things I've built, measured, and broken.
- The training machine gave up firstThe experiments still had questions left. The machine carrying them became unstable under full load, so I retired the whole system without pretending I had diagnosed the failed component.
- Every Tuesday, 199 countriesThe visa library contains entries for 199 countries and regions. Every Tuesday, automation checks the pages it has been configured to watch; people decide whether a change means anything for policy.
- Three models, one scoreI fine-tuned two 8B models, then tried a commercial model. All three scored 0.923. The problem was not the models. It was the evaluation.
What I am doing now
Now
- 2026.08 — Continuing to improve Piaopiao Travel from the networks where people actually use it.
- 2026.08 — Organizing state, recovery, and handover notes for systems meant to keep running.
- 2026.08 — Preparing to write down more of what I have built, measured, and broken.
Say hello
Email is the best way to reach me.
Send a note if something here made you curious, or if you want to compare notes about building and operating things.
If you're here for the professional version—what I can do for an organization—that lives at jyye.org