Nico Purnomo
Product manager. Data, ML, and the occasional rabbit hole.
I build data, ML, and AI products professionally, and share the analytical side here — interactive projects, articles, and ideas I find interesting.
↗ terminal-website-npurnomo.vercel.appLatest
Quorum — Cost per Correct Decision
An eval-governed control plane for visual AI decisions, built deliberately on a weak detector to prove the system matters more than the model. Confidence routing across a cheap detector, a vision-language model, and a human queue — every decision priced and recorded. The eval harness caught the detector naming 3 of 514 objects correctly, watched a drift alarm fire on cue, and picked up a never-trained class zero-shot. Includes an animated explainer with a locally-synthesized voiceover.
Work
World Cup Pub: 0→1 in 48 Hours
From idea to live app in 48 hours — spec, Jira backlog, React frontend, Maps integration, deploy. Post-mortem with the bugs left in.
World Cup Pub, Part 2: Steering the Agent
The human side of a 48-hour AI-assisted build — what I caught, missed, and would do differently. Covers failure taxonomy, prompt discipline, and a repeatable eval checklist.
Industry Financials
Interactive dashboard comparing financial metrics across US industries. Pick any two metrics and explore the relationships.
Remember Covid?
Excess death estimates by country using The Economist's model. Confidence bands show estimation uncertainty across 2020–2024.
Earthquakes & Extreme Values
Global seismic data mapped and modelled with Generalised Pareto Distribution. Estimates the probability of extreme magnitude events.
SQL Away!
Write SQL queries live in the browser against a customer orders dataset. Powered by DuckDB-WASM — no server required.
Heathrow Forecast
ARIMA model forecasting passenger traffic at Heathrow. Dummy variable accounts for Covid — with and without comparison.
Up Banking & Neobanks
How digital-only banks navigated the pandemic, what Xinja's collapse meant for the industry, and why Up's model was different.
Simile, Here & Everywhere
On similarity as a concept — from cosine distance to clustering to recommendation systems. Why measuring likeness is harder than it looks.
Data Reduction through Dynamic Range
Audio compression, photography, and SVD — a unified lens on what it means to reduce complexity while preserving meaning.