The Measurement Was Harder Than the Model
I distilled a 1.7B model into a BLS tool-calling agent. Six measurement mistakes in a row, a val-loss trap that was a config bug, a fix that came from changi...
AI Engineer at FUSE. Macro Data Scientist at JGP. MIT MicroMasters. Writing about agents, markets, and the code between them.
Recent posts
Winning the Google Cloud AI Hackathon: Building Multimodal Medical Annotation at Scale
How we designed a production-oriented multimodal system for medical image annotation and won the December 2025 Google Cloud AI Hackathon.
Analysing the impact of AI jobs at World Bank DataDive 2025
Our team built an interactive dashboard analyzing global AI job market trends and skills gaps across 66 countries using Stanford HAI data and World Bank labo...
MCP: Inside the Data Layer and the Choice of JSON-RPC 2.0
technical reflection on how the MCP structures agentic communication
Architects, Anthills, and AI: A Nobel Prize’s Lesson on Scientific Progress
chemistry nobel prize offers a profound lesson on scientific progress, with the AI acceleration
Spotify2YouTube is Back!
now I have a youtube API quota to construct the database
How to Add a Favicon to Your Jekyll Site
favicon is a small icon that appears next to your site’s title in browser tabs
Volatility Smile Explorer
an interactive dashboard for visualizing the volatility smile, calculating option prices with black-scholes, and exploring implied volatility across differen...
Why Lack of Correlation Doesn’t Mean Independence
pearson’s ρ drops to zero for X vs X —a neat example of hidden dependence.