Thoughts, systems & orthography.
Documenting everything in the open: daily learning logs across ML & languages, project devlogs, architectural breakdowns, and bite-sized TIL notes.
Starting from Zero: Why I'm Learning Japanese and Mandarin Simultaneously
Exploring the cognitive synthesis of dual-orthography immersion, shared kanji/hanzi semantic anchors, and engineering a bidirectional spaced repetition engine for dual language fluency from day one.
Sun, day, everyday diligence. Shared identical stroke sequence across both writing systems anchoring 420+ downstream compounds.
All Logs & Field Notes
6 itemsNo field notes match your filter.
Try searching for other keywords or select "All Posts".
The 365-Day Japanese Journey: Day 1 Kickoff
Setting up the daily 90-minute routine, phonetic kana muscle memory, and configuring Anki decks powered by stroke-order mnemonics.
Daily TIL: LoRA Rank Trade-offs & FlashAttention Intuitions
Why scaling rank r from 8 to 64 yields diminishing returns on stylistic alignment, but dramatic spikes in memory overhead.
r=16, alpha=32: sweet spot for attention query-key matrices
Chinese Tones Contour Intuition: Singing Pitch Analogy
Treating the four tones as musical vocal warm-up intervals: First (high steady soprano), Second (questioning rise), Third (low dip), Fourth (sharp command).
Building PolyglotLens: From Idea to Working Vision-OCR Model
An end-to-end devlog detailing dataset distillation for historical vertical Japanese scripts and dense handwritten Chinese characters.
Architecting HanziFlow: Fast Local Radicals Storage with SQLite
Storing 214 Kangxi radicals, stroke decompositions, and audio waveforms offline in browser WASM threads.
Demystifying Small Language Models: What Actually Happens During Fine-Tuning
Peeling back the math of rank-decomposition matrices, low-rank adaptations, gradient flow across frozen weights, and testing local inferences with 1B-3B models.