workCDSwritingsoonauditpromoter aboutori

Work

pS1RE · PROJECTS SINCE 20251,925 commits

pS1RE

Projects: 5
Written up: 3
In progress: 2
Updated: 2026-10-15

01nf-audit

02Proteus

03Press1POS

04adora

05thalweg

See the binder

01 nf-audit62–68%of a clean nf-core/rnaseq run's compute sits idle
02 Proteus3.9×better than chance at finding designs that bound
03 Press1POSLivein a US retail store since August 2026
04 adorain progress · no write-up yet
05 thalwegin progress · no write-up yet

Pick a tab · drag the binder to turn it

Close Esc

Project 01Released · v0.2.0

nf-audit

Shows where a Nextflow pipeline's cloud bill goes, task by task, and writes the settings that stop it paying for compute nobody uses.

62–68%of the compute paid for in a clean nf-core/rnaseq test run sits idle¹

public pipeline runs checked²: 61
current release³: v0.2.0
Role: Author
Stack: Rust
Source: github.com/OtoYuki/nf-audit

Sources

nf-audit, docs/rnaseq-61-runs.md: 28 of 30 releases
nf-audit README, Status
nf-audit releases, 15 Oct 2026

Price your own pipeline ➔

$ nf-audit analyze --trace … --report …

nf-core/rnaseq 3.15.1 · AWS test run

metered$79.23

never used$52.08 · 66%

QUALIMAPcpu 16% · mem 14%

wrote nf-audit.config

QUALIMAP6 → 2 cpus · 36 → 7 GB

nf-audit analyze · rnaseq 3.15.1

The problem

Every Nextflow run records what each task asked for and used. Nobody turns that into money, so a team sees the cloud bill, not which step ran it up.

What it does

Prices every task from those records, shows what was paid for and never used, and writes a config that asks for less. Offline, on any cloud or cluster.

Proof

Totals match the pipeline’s own report on all 61 public nf-core/rnaseq runs checked.
Found a step asking for 72 GB and using under 10. The nf-core maintainer: “Your numbers check out”. Fix in review.
A docs fix on how Nextflow measures task time, merged into Nextflow.

Close Esc

Project 02In development · v0.9.1

Proteus

Ranks predicted protein designs, so a lab spends its GPU hours and bench time on the ones most likely to work.

3.9×better than chance at finding designs that bound in the lab¹

structures checked against reference tools²: 53
current release³: v0.9.1
Role: Author
Stack: Rust
Source: github.com/OtoYuki/proteus

Sources

proteus README; 3,669 designs, Overath et al. 2025
proteus, validate/corpus.toml
proteus releases, 15 Oct 2026

Start a project ➔

$ proteus analyze designs/ --interface A:B

12 designs against IL-7Rα · lab results hidden from it

rank · designscore · bound in the lab

1 il7ra_binder_af2_480.698 · yes

2 il7ra_binder_af2_340.646 · no

3 il7ra_binder_af2_930.622 · no

4 il7ra_binder_af2_940.587 · no

… 8 more, down to 0.000

proteus analyze · 12 of 3,669 lab-tested designs

The problem

A design campaign ends with thousands of predicted structures. Picking which to make usually means a pile of scripts and numbers nobody has checked.

What it does

One program measures every model in a folder and ranks the designs. Every measurement is checked against an established tool.

Proof

On 3,669 designs tested in the lab, it finds the ones that bound 3.9× better than chance.
53 structures re-measured against reference tools on every change.
A fix to how OpenFold3 seeds predictions, merged upstream.

Close Esc

Project 03In production

Press1POS

A point of sale that keeps selling when the internet drops. I lead it end to end, from the countertop terminal to the cloud.

Livein a US retail store since August 2026¹

checkout exchanges replayed on every change²: 466
current version³: v0.2.16
Role: Technical Director, press1-inc
Stack: Rust · React · Android

Sources

press1-inc (the store is kept private)
press1-pos README, golden master (private repo)
press1-pos README, 15 Oct 2026 (private repo)

Start a project ➔

The Press1POS home screen: twelve areas, among them the register, cash drawer, orders, reporting, inventory (58 out of stock), customers, vendors, purchase orders ($8,850.28 on order), restocking, a customer-facing screen and setup. — home screen · v0.2.16 · the countertop terminal

The problem

A shop’s till can’t stop when the internet does.

What it does

Sells, prints receipts, tracks stock and closes the till the same way with or without a connection, then syncs when it’s back. One Rust core runs on the countertop terminal and the desktop.

Proof

466 recorded checkout exchanges are replayed against a fresh server on every change.
Money is kept in whole cents, never floating point, with the tax rounding written down and tested.