Hi Product Hunt! I built BitLens because quantizing neural networks for microcontrollers and edge hardware usually feels like flying blind. When you squeeze a PyTorch or ONNX model onto an STM32H7 or edge NPU, standard post-training quantization treats the model like a black box. You quantize to INT8 or INT4, accuracy drops, or the tensor arena throws an out of memory error at runtime, with no visibility into which layer caused it. BitLens gives you a visual DevTools experience for tensors: 1. Python Profiler: Captures 128-bin histograms, activation outliers (Omega = absmax / p99.9), kurtosis, simulated SQNR, and topological peak SRAM liveness in one calibration pass. 2. Web Studio: A zero-backend Next.js inspector preloaded with MobileNetV3, ResNet-18, and ViT-Tiny. Click any layer to see quantization grid steps overlaid on weight and activation distributions. 3. Greedy Pareto Auto-Fit: Ranks layers by SQNR loss per byte saved to fit your target memory ceiling. 4. 1-Click Code Exporter: Generates ready-to-run scripts for PyTorch torchao, ONNX Runtime, and TFLite Micro. It is 100% open source under the MIT license. I would love feedback from fellow developers and ML engineers on the outlier heuristics and exporters!

BitLens

Chrome DevTools for neural network quantization & MCU memory

Chrome DevTools for neural network quantization & MCU memory