Gridbert
Started: Feb 25, 2026
Energy agent for Austrian consumers. Consumption analysis, tariff comparison, anomaly detection.
LinkedIn Post Jakob Kreisel 
Beispielbild von Jakob 
Screenshot Report 
Gridbert Screenshot After one day, the first prototype is standing. Give it a bill and smart meter credentials, it fetches tariff switching info and checks whether joining an energy community makes sense. All local with Ollama, no AI calls to the cloud. Bummer: the agent loop got cut because the first step is very deterministic. Deployment without SaaS isn't trivial either. And honestly: right now it's an error-prone E-Control wrapper for people who can't read PDF bills. Need to let it rest and get feedback.
Ollama with Qwen2.5 7B locally — billing data shouldn't go to the cloud. Trade-off: 7B models can't do reliable tool calling. Llama3.1 output tool calls as plaintext, Qwen2.5 marginally better but not stable. Consequence: agent loop dead, replaced by hard-wired 5-step pipeline. LLM only extracts JSON from bill text (with two-turn hack), rest is deterministic Python. E-Control as tariff API regularly doesn't respond within 30 seconds — retry with backoff doesn't fix that the API isn't designed for machine queries.