Stylus Trajectory CNN

Lightweight CNN for stylus trajectory regression on panel-IC hardware. Inventec Corporation, Digital Center (AI-on-Chip team), 2023 — sole ML engineer on this sub-project.

A convolution-based regression model that turns noisy capacitive-sensor readings into a stylus trajectory, replacing the existing rule-based on-chip algorithm. Deployed in the panel IC of a top IC design house. Sub-2-pixel accuracy at 10–50K parameters (depending on hardware variant), ~50% error reduction over the prior rule-based baseline that had been hand-tuned for years by domain experts.

What this project actually shows.

• Lightweight ML targeting embedded hardware. The 10–50K-parameter budget was set by the panel IC’s compute and memory envelope; the model was designed working backward from that constraint, not by shrinking a research model after the fact.
• Beating a tuned rule-based baseline in production. ~50% error reduction over a domain-expert-tuned baseline is a meaningful applied-ML win — these baselines are usually hard to beat because they encode years of domain knowledge.
• End-to-end ownership of a shipped product feature. Dataset construction from real sensor traces → model design under hardware constraints → training → evaluation against the production baseline → handoff for customer-IC integration.

Stack. PyTorch for model design and training; a regression pipeline targeted at the pixel-accuracy metric the customer cared about; evaluation harness for apples-to-apples comparison against the on-chip rule-based algorithm.

Context. This project ran out of Inventec’s Digital Center (AI-on-Chip team), in parallel with the ICASSP 2024 paper on cross-domain augmentation for foot-ulcer segmentation that ran out of Inventec’s AI Center. Customer name and customer-product details are confidential.