Cinderella Local AI PHASE 1 INTERIM REPORT We have successfully brought Cinderella's first local AI inference system online. Hardware Windows 11 workstation NVIDIA GeForce RTX 3090 with 24 GB VRAM Intel Core i9-11900 128 GB system RAM Runtime Ollama 0.32.3 running natively on Windows Qwen 3.5 9B, Q4_K_M quantisation OpenAI-compatible HTTP API German and English intent classification Strict schema-conforming JSON output Reasoning output disabled for deterministic classification Privacy and security The inference API listens only on 127.0.0.1:11434. It is not exposed to the LAN or the public internet. Ollama Cloud and Ollama Web Search have been explicitly disabled. Local prompts and archived member content remain on operator-controlled hardware. External providers will only be available through explicit, auditable configuration. Safety boundary The AI only classifies messages into Cinderella's closed intent catalog: PUBLISH UNPUBLISH STATUS SEARCH PRICE HELP UNDO UNKNOWN The model never executes actions. Cinderella's existing deterministic rule engine remains active as the automatic fallback. Consent, permissions, confirmation prompts and all actual actions remain controlled by the existing application code. Accuracy 6 of 6 bilingual classification tests passed. German and English were detected correctly. All returned intents matched the expected catalog entry. Structured output remained valid JSON with no invented fields. Measured performance Warm request latency: approximately 0.7 to 1.1 seconds Time to first token: 759 milliseconds Complete validated JSON response: 1.075 seconds Cold start latency: 6.636 seconds Cold model loading time: 6.009 seconds Generation speed: approximately 65 to 73 tokens per second Concurrency Four simultaneous requests completed in 2.394 seconds. Measured parallel throughput: 1.67 requests per second. All concurrent classifications were correct. GPU usage The model runs entirely on the RTX 3090. Loaded model size: approximately 6.7 GB VRAM remaining after loading: approximately 15 GB Observed temperature under load: 62 degrees Celsius Observed power usage under load: approximately 284 watts Reliability Ollama starts automatically after normal Windows login. The API remained available after a complete Windows reboot. After a forced process termination, Ollama restarted automatically in approximately two seconds. The model is unloaded from VRAM when idle and loaded again when required. Planned endpoint modes Local only OpenAI Claude Local with explicitly approved cloud fallback Automatic workload routing Provider comparison mode External transmission will never happen silently. Archived member content and moderator-private material will remain local by default. Next steps Sustained load measurement Vision model evaluation Personality and conversation model evaluation Secure private connection between the Windows GPU machine and the Debian VPS Authenticated Cinderella AI Gateway Administration, statistics, provider selection and cost monitoring