Local LLM

GelbIT and FTC Dashboard both use local LLMs via Ollama. For GelbIT, I fine-tuned Gemma 3 on German recycling rules, reaching 94% classification accuracy. Running models locally means no API costs, no rate limits, no data leaving the machine, and sub-second response times on consumer hardware.

import ollama

response = ollama.generate(
    model="gemma3",
    prompt="Classify this waste item: empty yogurt cup",
    options={"temperature": 0.1},
)

print(response["response"])
# → yellow bin (Gelber Sack)