The author faced a common data science headache: thousands of short free-text annotations that all meant similar things but used completely different wording. One person wrote "test code, not deployed anywhere" while another wrote "only runs in CI/CD pipeline during integration tests" - traditional clustering couldn't handle the paraphrase variation.

The solution was using a locally hosted LLM via Ollama as a zero-shot classifier. Instead of asking an algorithm to discover clusters, the author defined candidate categories upfront and asked the model to classify each entry. Research from Chaeand Davidson (2025) and Wang et al. (2023) shows LLMs in zero-shot mode can match or outperform fine-tuned models on classification tasks.

For ~7,000 entries, the pipeline ran Gemma 2 (9B parameters) on a MacBook Pro in about 45 minutes. Preprocessing cut the token budget by roughly 30%, and temperature at 0.1 kept classifications consistent. The results were striking: over a quarter described non-production environment issues, while 21.9% were risks already handled by security frameworks. This approach works best for medium-scale datasets with semantically complex text where you can define categories but lack labeled training data.