A compiler for vision models: describe the detector you want in plain English, get a deployed TensorRT engine, or a documented refusal, not a silent degradation.
Public repo (MIT)Prompt2Model takes a plain-English description of a detector you want and runs it through a full pipeline: prompt parsing, dataset loading (classification and COCO-style detection), training, ONNX export, and evaluation.
An opt-in factory-compiler layer adds an LLM planner, a distill-then-quantize-then-accuracy-floor compression gate, pluggable deployment targets, calibration and conformal abstain behavior, and a flywheel hard-case store.
The classification path is described in its own README as fully validated end to end. The detection path is integrated through the project's week-4 milestone but is not yet fully validated at every stage, and the README says so directly.
There is no accuracy or latency benchmark table in the README yet. The demonstrated mechanism is the compression refusal gate: it refuses to ship a compressed artifact that falls below max(accuracy_floor_relative * baseline_accuracy, constraints.accuracy_floor) (default relative floor 0.98), and keeps the uncompressed model instead. This is a behavioral guarantee, not a measured accuracy number. The README states the classification path runs end to end on synthetic, generated toy data.
Measured 2026-07-12 in a fresh venv: 118 passed, 0 errors, 0 failures. Previously cited inconsistently between 115 and 117 across sources; that gap is now closed to one canonical count.
The repo is public and MIT-licensed, so an outsider can clone it, run the classification path end to end on the included synthetic toy data, and inspect the compression gate's refusal behavior directly. What cannot be reproduced today: a real-image accuracy benchmark, because none has been published yet, and full validation of the detection path, which the project's own README flags as not yet complete.