Welcome — what this blog is about

This is where I write about the two halves of my work:

Backend & systems — production NestJS/Postgres services, plus side-experiments in Go (e.g. Tokimeki, a masterless filesystem-based job runner).
AI research — open-set recognition, OOD detection, multi-label classification, and dynamic graph embedding.

Posts in the pipeline

Designing Tokimeki — why no broker, why the FS is the queue, and the tradeoffs.
OSR metrics, demystified — what AUROC actually measures on open-set tasks, and why it’s misleading.
Bayes-optimal multi-label inference — the math behind inference_probabilistic_mlc.

Math renders fine, inline: $\hat{y} = \arg\max_{y \in \mathcal{Y}} P(y \mid x)$ , and as a block:

\mathrm{AUROC}_{\text{open}} = \int_{0}^{1} \mathrm{TPR}_{\text{ID}}(t) \cdot \mathrm{FPR}'_{\text{OOD}}(t) \, dt

Mermaid diagrams too:

flowchart LR
    P[pending/job-001.json] -->|atomic rename| O[owned/A/job-001.json]
    O -->|worker A finishes| D[done/job-001.json]
    O -->|heartbeat stale| P

Code blocks too:

func (r *Runner) Claim(job Job) error {
    // atomic rename = atomic claim, no broker required
    return os.Rename(job.PendingPath(), job.OwnedPath(r.ID))
}

Stay tuned.