Daytona HackSprint Seoul · 2026-09-19
InBody Auditor
The numbers are free. The story is the product.
Ten clinical body scans. Two years. One page that finally says whether it is working.
Daytona × Nosana × DNSimple.
The problem
I had every number. They never told me if I was okay.
10clinical scans
14metrics each
140numbers total
- A scan prints a sheet. A sheet is not a story.
- I could not tell whether I was losing fat or muscle without doing the maths myself.
- Progress you cannot read is progress you cannot keep.
The real story · my data, all ten scans
Up, then down. The chart I could never see.
The year I lost
Then I stopped, and the sheet went red.
+19.1kg gained
+21.8kg pure fat
−2.1kg muscle lost
- Visceral fat doubled: 119.8 → 235.4 cm².
- InBody score crashed: 69 → 44.
- A year of sitting cost me more than the eleven months before it.
The claw back · Jun to Sep 2026
Five months later: fat down, muscle kept.
−6.7kg total
−6.4kg pure fat
+0.2kg muscle held
- 95.5% of the weight I lost was pure fat. Typical GLP-1 loss without lifting gives up 25–40% of it as muscle.
- Visceral fat down 17.2 cm². InBody score back up to 50.
- I only know any of this because I did the maths by hand. That is the bug.
What this proves
Nobody reads 140 numbers. That is the whole product.
- The scan gives you raw clinical data — weight, SMM, fat mass, PBF, visceral fat, score.
- What you need is one verdict: am I losing fat, or am I losing muscle?
- I did that reading by hand for a year. It should be a page.
The scan is data. The page is the decision.
The market
Nobody wants another tracker.
- 17 lifting-positioned apps in the store, every one at 0–8 ratings.
- 602 reviews read. Half the complaints are not about features.
- 51% of negative reviews are paywalls — "free app but pay to use", "bait and switch".
- The pain is not missing features. It is that my own story sits behind a subscription.
So this is not a tracker. It is a translator.
The idea
Numbers in.
A story out.
Upload your scan history. Get back a plain verdict anyone can read:
how much you lost, whether you kept muscle, and your daily protein floor.
Model writes one coaching line. Page gets a fingerprint. One link, no subscription.
Live · the app's own output page
Chart, verdicts, protein floor, and a coaching line from qwen/qwen3.8-27b on Nosana. Served from a Daytona sandbox by signed URL.
How it works · three sponsors, all load-bearing
Daytona
A fresh sandbox runs the maths and is destroyed after. I can hand it a file I do not trust.
Nosana
Runs the model that writes the coaching line. Live model id resolved at runtime.
DNSimple
Publishes a fingerprint of the finished page to DNS. That is the receipt.
- 15-minute sandbox TTL, destroyed with
wait=True. - Isolated on the way in, fingerprinted on the way out.
Why me · every number checkable
I have already shipped to 100,000 people.
Spiko — AI English Exam Prep
100,000+ installs on Google Play
4.6 stars, 77 ratings
Cash-positive after infrastructure, inference and platform fees
play.google.com/store/apps/details?id=org.milliytechnology.spiko
verified on the live listing, 2026-09-19
- That app won where the store still worked. This category is the opposite: 17 lifting apps, all at 0–8 ratings.
- And this is not a demo. Ten scans, two years, and a written plan to 68 kg — running on my own body.
Close
The numbers are free.
The story is the product.
Thank you.
Appendix · for the code-level review
Where each sponsor lives in the source
auditor.py · Daytona: create(), set_ttl(15), code_run(), delete(wait=True)serve.py · Daytona: serves the report with create_signed_preview_url(3000, 900)coach.py · Nosana: GET /v1/models to pick the model, then one chat completionpublish.py · DNSimple: POST /zones/<zone>/records with the sha1 of the reportanalyze.py · stdlib only, the deterministic part
