Tiny-CNN object detection · Apache-2.0
tensor-factory turns a text prompt into a synthetic dataset, an int8 ONNX model, and a CPU inference harness. No GPU at runtime. No AGPL anywhere.
Fig. 01 — Reference numbers · helicoils
A detection is four uint8 values — normalized xyxy, one byte each.
Fig. 02 — The pipeline
Generate images from a prompt — hosted Gemini, or a mock that runs anywhere.
GroundingDINO proposes boxes; a human gate approves them before they train.
A tiny soft-argmax CNN, quantized down to an 81 KB int8 ONNX model.
onnxruntime on CPU — driven from a CLI, an MCP server, or an HTTP endpoint.
Point it at a new object and turn the crank — the pipeline, in full →
Fig. 03 — The documentation
Clone, sync, and run the bundled int8 detector on CPU in under a minute. No GPU, no API keys.
The five steps that turn a prompt into your own detector — synthesize, label, review, train, run.
Run the helicoil detector in your browser, entirely client-side.
Every command and flag across the five packages, plus the install extras and env vars.
BBox geometry, the 8-bit codec, the Detector, the review gate, and the training functions.
Why a detection is four bytes, the soft-argmax model and its learnable gain, int8 quantization, and the review gate.
Fig. 04 — Examples
Locating coiled-wire threaded inserts in machined parts from low-cost microscope imagery — what every package is currently tuned against, and the template for the next object you point the factory at.
Try the live demo →