Phase 03 — From Models to Browser Inference

This phase moves from hand-built kernels to real exported models. Success now depends on three contracts agreeing: model graph and weights, runtime operator/backend support, and the media transformations surrounding inference.

Branches

• Model Formats, Conversion & Quantization — ONNX, TFLite/LiteRT, safetensors, specialized formats, graph optimization, shapes, operator coverage, precision, calibration, and numerical drift.
• Browser Inference Runtimes — ONNX Runtime Web, LiteRT.js, Transformers.js, MediaPipe, TensorFlow.js where relevant, sessions, tensors, lifecycle, execution providers, and fallbacks.
• Preprocessing, Postprocessing & Media Pipelines — image/audio/text decoding, resize, normalization, layouts, tokenization, masks, thresholds, and deterministic output interpretation.

Workbench milestone

v3 — Real models and real runtimes: export the same small model to ONNX and TFLite/LiteRT in FP32 and a justified quantized form, execute it through ONNX Runtime Web and LiteRT.js, and compare each complete pipeline with golden outputs.

What this phase establishes

• Reproducible export commands and artifact manifests.
• Operator and shape compatibility checked before product integration.
• One conceptual backend interface without erasing runtime-specific behavior.
• Explicit tensor ownership, disposal, preprocessing, and postprocessing.
• Numerical comparisons that separate conversion, quantization, runtime, and pipeline drift.

Exit criteria

• Every model artifact has source, format, precision, checksum, size, and operator assumptions.
• The same fixtures run through both runtime paths or produce an explicit unsupported result.
• Preprocessing and postprocessing match the reference implementation within declared tolerances.
• Initialization, first inference, warm inference, peak memory, cleanup, and fallback are recorded.

Continue with Phase 04: Hardware Acceleration.