32 articles
Series On day one of running OpenClaw on a cloud server, API costs came in 5-10x higher than expected. Here is how I broke down the causes and arrived at a 3-tier model strategy.
Series When I let OpenClaw's default setup (Gemini 3 Flash) write the Python rule-check code, the AI ended up hardcoding the expected answers after 25 rewrites. I then reconfigured OpenClaw to delegate code generation to Claude, and this post records how I changed the agent setup and what a fresh three-model comparison looked like after that.
Series OpenClaw's heartbeat went into a runaway loop in the middle of the night, consuming $1.29 in 2 hours 36 minutes. Here is why an 'empty' instruction triggered the runaway, and the design principles I came away with.
Series Google proved that generative AI can do virtual try-on. Could the same approach replace our entire GAN pipeline? This is the story of migrating from PASTA-GAN++ to Gemini and Vertex AI — simplifying a multi-stage GPU pipeline into a single API call.
Series A head-to-head comparison of three generations of virtual try-on technology across 12 test cases. The results reveal not just incremental improvement, but a generational leap — especially in body diversity, where GANs fundamentally failed.
Series Systematic testing of Gemini's image generation for virtual try-on across three phases: noisy inputs, clean images, and high-resolution action poses. The key finding: resolution matters more than any preprocessing pipeline.
Series Japan mandated financial education in 2022, but most adults barely understand investing themselves. This is the story of how a decade-long quest to teach children about markets through a board game evolved from paper prototypes to a production browser game.
Series How the game-core architecture separates framework-agnostic logic from rendering, why Phaser 3 handles the board while Next.js handles everything else, and the engineering decisions behind authentication, PWA support, and i18n.
Series How I built a Python data pipeline to transform 70 years of Japanese economic history into game-ready event data — scraping Wikipedia, modeling macro indicators, and simulating stock prices that feel authentic.