Groundtrack: Graded Anomaly Detection
AI Builders Challenge with IBM Bob

Case study

AI Systems / Agents

Competition entry

Dhaka, Bangladesh

AI Systems / Agents · Detailed breakdown

Project overview

Why this build
mattered.

IBM Bob wrote a spacecraft telemetry fault detector. The grader was locked first and 26 of 81 NASA recordings were held out, so every improvement is scored on data the agent never saw.

Snapshot

AI Builders Challenge with IBM Bob / Competition entry / Python, IBM Bob, IBM Granite

PythonIBM BobIBM GraniteVercel
0.266 to 0.623
heldOutF1
128 to 7
falseAlarms
0.163 to 0.731
precision

System views

Screens and
surfaces

What shipped inside

Key features
and modules

The architecture, workflow and results that made this project useful.

Highlight

Held-out F1 rose from 0.266 to 0.623 and false alarms fell from 128 to 7

Highlight

Eight rounds were tried and the grader kept one

Highlight

IBM Granite writes the operator briefs; a live console shows the results

Delivery summary

Build notes

Project essentials

Groundtrack: Graded Anomaly Detection

AI Builders Challenge with IBM Bob · Competition entry · AI Systems / Agents

Held-out F1 rose from 0.266 to 0.623 and false alarms fell from 128 to 71
Eight rounds were tried and the grader kept one2
IBM Granite writes the operator briefs; a live console shows the results3

Let's build something worth building.