The agentic approach
The answer is a process, not a completion.
Phidox does not choose one canned query and decorate the result. It explores, changes course, writes analysis code, and keeps a compact memory of what the evidence established.
One question
many possible paths
one evidence-backed answer
A visible working loop
Curiosity, with an audit trail.
Every database exploration, model call, and generated script appears as a collapsed step with its duration. Open any step to inspect the SQL or code.
Orient
Read the data catalog, inspect relevant tables, and decide what evidence the question actually needs.
Explore
Run small SQL probes and schema checks. Let the results change the next move.
Build
Write a purpose-fit Python or JavaScript analysis when a query alone is not enough.
Test
Check cohorts, baselines, sample sizes, dates, and model limitations before drawing a conclusion.
Present
Turn the evidence into a compact interactive result, with the work log still attached.
Working memory
Long enough to think. Small enough to stay sharp.
Tool results enter the live context as the analyst works. When raw outputs become bulky, older evidence is compacted while the newest observations stay intact. The model can keep exploring without dragging every table through every subsequent step.
At the end of a turn, Phidox stores a bounded analytical memory—not the rendered HTML, chart SVG, or every intermediate row. A follow-up starts with what was established and can re-query live data when the question changes.
Read-only DuckDB over verified cloud snapshots.
Purpose-built scripts for statistics, transforms, and visual evidence.
Three measured signals, called only when they improve the answer.
A complete, sandboxed result card that can be shared with the community.
Give it something difficult