<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Kevin Yuan</title><description>Building at the intersection of software, hardware, and machine learning.</description><link>https://raviolico.de/</link><language>en-us</language><item><title>Byte-Level Isn&apos;t Dead: Adding a Classification Head to BLT-1B</title><link>https://raviolico.de/research/notes/blt-classification-head/</link><guid isPermaLink="true">https://raviolico.de/research/notes/blt-classification-head/</guid><description>Last time, we concluded byte-level models were not a good fit for clinical text classification. We now think the conclusion was overgeneralised. This note describes a first attempt at adapting Meta&apos;s Byte Latent Transformer for classification, with preliminary results on SST-2.</description><pubDate>Wed, 29 Apr 2026 00:00:00 GMT</pubDate></item><item><title>The Nikon Coolscan, Re-implemented</title><link>https://raviolico.de/projects/nikon-coolscan-reimplemented/</link><guid isPermaLink="true">https://raviolico.de/projects/nikon-coolscan-reimplemented/</guid><description>Reverse engineering an aging film scanner&apos;s firmware, drivers, and SCSI protocol.  Then building a clean-room H8/3003 emulator that NikonScan can&apos;t tell apart from the real thing.</description><pubDate>Mon, 20 Apr 2026 00:00:00 GMT</pubDate></item><item><title>Reverse Engineering the Alientek HP20 Reflow Hot Plate</title><link>https://raviolico.de/projects/alientek-hp20-reverse-engineering/</link><guid isPermaLink="true">https://raviolico.de/projects/alientek-hp20-reverse-engineering/</guid><description>Cracking XOR encryption, decompiling .NET updaters, and reverse engineering BLE protocols on a $100 soldering hot plate.</description><pubDate>Wed, 25 Mar 2026 00:00:00 GMT</pubDate></item><item><title>Bigger Isn&apos;t Always Better: Modern LLMs vs. BERT for Clinical Text Classification</title><link>https://raviolico.de/research/notes/llms-vs-bert-clinical-nlp/</link><guid isPermaLink="true">https://raviolico.de/research/notes/llms-vs-bert-clinical-nlp/</guid><description>We benchmarked six modern architectures against our published Bio+ClinicalBERT baseline for classifying antibiotic indications from EHR free-text. On this task, a 110M-parameter BERT matched or came within one F1 point of models 245x its size. The pattern is consistent with domain pretraining mattering more than scale, but the comparison comes with caveats.</description><pubDate>Thu, 12 Mar 2026 00:00:00 GMT</pubDate></item><item><title>ChipWhisperer Lite: Self-Assembly with USB-C</title><link>https://raviolico.de/projects/chipwhisperer-lite-self-assembly-usbc/</link><guid isPermaLink="true">https://raviolico.de/projects/chipwhisperer-lite-self-assembly-usbc/</guid><description>Building a ChipWhisperer Lite from scratch in 2025: current BOM, USB-C in place of Mini-USB, Gerbers on GitHub.</description><pubDate>Mon, 08 Sep 2025 00:00:00 GMT</pubDate></item><item><title>Tufte on the Web · A Typographic Swatchbook</title><link>https://raviolico.de/projects/zz-test-elements/</link><guid isPermaLink="true">https://raviolico.de/projects/zz-test-elements/</guid><description>A layout system I built for this site, adapted from Tufte&apos;s handout style and rewritten for the web with modern touches like responsive margin notes, collapsible sidenotes on mobile, and full-bleed figures.</description><pubDate>Wed, 01 Jan 2020 00:00:00 GMT</pubDate></item></channel></rss>