👻 The Séance Engine: Finding the Ghost in the Data with TabPFN & Gemma

Description and use

A Python automation script reads a day of system logs, uses TabPFN to flag the most anomalous entry, and prompts local Gemma to generate a ghostly narrative, all running locally.

Problem

Detect anomalous system log entries on a personal machine without sending data externally.

Outcome

The pipeline identifies a 3 AM log entry with 98% CPU as anomalous and produces a whimsical narrative about a ghostly process, all running locally in ~1 min.

Workflow

• Load CSV of 24 hourly log rows
• For each row, perform leave‑one‑out evaluation with TabPFN: train on 23 real rows + noise, predict p_real for held‑out row
• Identify row with lowest p_real as anomaly
• Prompt local Gemma model via Ollama to generate a ghostly story incorporating the anomaly details
• Output story to terminal and record trace in Sentry
• Optional: run with SENTRY_DSN to capture metrics

Evidence

Everything runs on my laptop: one Python file, about 100 lines.

Source

The anomalous row, including its p_real score, goes into a prompt that asks Gemma to speak as a gossipy, theatrical ghost…

Source

The trace showed me something I didn’t expect: the leave-one-out TabPFN scan (24 small fits on CPU) took 38 seconds, and Gemma took about a minute.

Source