← projects

# Astro

engineer, consultant · aug – sep 2026

Lost Astronaut is a venture studio in Vilnius that starts companies. Astro is its market-intelligence system: it pulls public market data into one database, looks for where money is moving, and puts the evidence in front of the partners for their weekly meeting. I took the job to see the other side: how a studio decides which companies to start, and what investors look for before they put money in.

## what i built

## what was hard

Measuring my own machine. The first version was 23 LLM agents running 17 analyst playbooks. When I measured it, it wasn't producing anything the partners could act on, so I rebuilt detection as statistics. Then I audited the replacement too: it was mostly detecting changes in our own data collection, with 2.7 false alarms per market-year against a budget of 1. The version that went live reads each week as it was known at the time, and a change-point detector (CUSUM) with thresholds calibrated by simulation flags real shifts. LLMs only read and judge what it flags.

## what i learned

Blank beats wrong. Scores should order things, never gate them. Measure on the production box before optimizing: a bigger server would have bought weeks, not a fix. Navigation should follow the questions people ask in the meeting, not the database tables.

## what happened

# this site is a shell: open it here