Vibe Coding Discoverrizzo-flow
View on GitHubThe open, local take on Jev: typed decisions from an LLM, without generating a single token
Local-first implementation of the Jev programing model: an HTTP API and CLI that read only the answer-letter logits from a single forward pass, returning typed, probability-bearing decisions (boolean, choice, score, number) from a GGUF LLM via llama.cpp with zero generated tokens.
Use Cases
Classify text into typed decisions with confidence scoresSupport ticket triage and urgency detectionRubric-based scoring and ratingNumeric estimation from unstructured stateRouting/decision logic in agent pipelinesLocal inference without sending data to a hosted APIDrop-in local replacement for the Jev system-one APIFast low-latency per-decision inference (e.g. game AI)
Built With
- Language
- Python
- Frameworks
- llama.cpp · FastAPI · Pydantic · Uvicorn · Jinja2 · MLX · MLX-LM · Hugging Face GGUF
Tags
local-llm · structured-output · logits · llama.cpp · typed-decisions · probability-calibration · classification · offline · gguf · quantization · fastapi · pydantic · inference · system-one · zero-generated-tokens · jev-compatible