← Index Case study / 02 2026

Fault-vs-noise classifier · 2026

Rogue Protocol

A fault-versus-noise classifier for ship logs, tuned to 0.6832 validation F1 and shipped through a checker-gated submission path.

Year: 2026
Role: Submission integrity, validation, Round 2 orchestration
Context: Team of 4
Stack: Python / pandas / scikit-learn / HGB + SVD-64 + TF-IDF

0.6832 validation F1

+0.081 over the LogReg floor

3,000 unlabelled logs classified

4 person team

01

Problem

A ship computer floods the crew with false alarms and hides real faults inside calm log lines.

02

Build

Shipped a classifier for 3,000 unlabelled ship logs; validation F1 0.6832 at threshold 0.54, +0.081 over the LogReg floor. Team result.

03

Decisions

Owned submission integrity as captain: threshold sign-off, checker-gated CSV, upload log, and final card.
Every upload passed TRANSMISSION VALID.
Led the Round 2 orchestration layer across triage, memory, scheduling, and notifications.
Generated handover and timeline reports from real traces, not hand-written summaries.

04

Outcome

3,000 unlabelled logs classified, with a checker-gated submission path from threshold to upload.

05

Links

Quiet Check-in

A local-first night-travel safety app that stays quiet while a trip is on track and wakes a contact only when a check-in is missed.