NeuroForge

NeuroForge X is a multimodal AI compiler and execution runtime, not a chatbot wrapper. You describe a multimodal workload declaratively (video, vision, reasoning, reporting); NeuroForge compiles it to a typed intermediate representation, type-checks it, optimizes the graph (dead-node elimination, common-subexpression elimination, parallelism analysis), and runs it on a dependency-aware concurrent scheduler with content-addressed caching, retries, timeouts and profiling. It runs entirely on local Meta Llama models: Llama 3.2 Vision performs visual perception on real video frames (decoded with ffmpeg), and Llama handles planning and reasoning, all via Ollama with no cloud and no API keys. The demo analyzes a warehouse video end to end: it detects the objects present, computes measured spatial features, reasons over them, and assembles a structured inspection report, explicitly separating measured observations from model inferences. Written in Go, around 5,000 lines, with 23 tests passing under the race detector.