Roadmap

What's done and what's next.

Crucible is active development, not a closed one-off. Below: everything already in the engine, what I'm working on now, and the on-device use cases the runtime is being aimed at.

Shipped

6 items

shipped

C++17 core with custom tensor class

Row-major flat memory, shape/stride bookkeeping, Eigen-backed matmul and conv2d.

shipped

Hand-written ONNX protobuf decoder

Reads .onnx binary wire format without protoc. Varint, length-delimited fields, embedded tensors.

shipped

Kahn's algorithm graph executor

Topological sort over the DAG, ready-set scheduling, deterministic execution order.

shipped

20-operator native engine, checked against ONNX Runtime

Conv, pooling, BatchNormalization, Gemm and activations — enough for ResNet18 and MobileNetV2 — with a 70-case parity suite against ONNX Runtime.

shipped

Small Rust WASM runtime

A separate pure-Rust interpreter (MatMul, Add, Relu, Sigmoid, Softmax) built with wasm-bindgen; runs the fraud model in the browser.

shipped

C ABI, Python module and HTTP server

Shared library with a C interface (used by the Rust CLI), a pybind11 module, and a FastAPI server with per-user model registry.

In progress

0 items

Planned

7 items

planned

Close the performance gap with ONNX Runtime

Single-threaded, Crucible is currently several times slower than ONNX Runtime on ResNet18 / MobileNetV2. Threading and faster conv paths are next.

planned

Native ↔ WASM parity harness

Automatically compare the two runtimes' outputs on the models they both support.

planned

Quantized int8 MatMul path

Symmetric per-tensor quantization for smaller models.

planned

Biometric matching (face embeddings)

On-device MobileFaceNet inference; cosine similarity in the tab, no biometric data uploaded.

planned

Document OCR (CRNN)

Client-side ID/receipt OCR — pixels never leave the device, only extracted fields cross the wire.

planned

WebGPU backend

Optional GPU kernels for the larger conv-heavy models; keep WASM as the always-available fallback.

planned

Streaming model loader

Range-request the .onnx from CDN and start executing before the full file has landed.