ST-1Signal Tracer

An analytics workbench
for humans and your AI.

No model bundled. ST-1 gives the AI you already use (Claude, Cursor, anything that speaks MCP) the tools it needs to do real analysis: DuckDB SQL, profiled columns, a curated visualisation catalogue, auditable lineage.

Chart example: bill length vs bill depth from the Palmer Penguins dataset, coloured by species.

For humans

A real workbench

Upload CSV or parquet. Profile every column. Cross-filter by clicking a histogram bar. Author derived tables in SQL with autocomplete and live preview. Build charts that honour your data's grain.

For your AI

Excellent at analysis and viz

No LLM is built in. ST-1 hands whichever model you trust a curated chart catalogue with when-to-use guidance, column profiling, data-quality flags, and DuckDB SQL. Your AI picks the right form, surfaces the bad rows, names the trade-offs.

For trust

Verifiable, not vouched for

SQL is canonical. Lineage is explicit. Every transformation an AI runs is a row you can read; every chart points at a real column. You don't take the model's word for it. You see what happened, and check it.

Scale

Always snappy. Considered at scale.

Three layers, sized to keep interaction in a single frame and the canonical data honest. The browser does what browsers do well; the server does what servers do well; the seam is explicit.

Up to ~2M rows

Sub-frame interactive

Mosaic precomputes column-summary cubes in DuckDB-WASM. Filter a histogram bar; the table, every other summary, and any chart on top all repaint in the same frame. No round-trips.

2M – 50M rows

Auto-sampled preview

Sources past the threshold materialize as a reservoir sample in your browser for snappy interaction. Every panel reading from the sample carries a clear badge. The canonical data is untouched on the server. Toggle off when you need exact counts.

50M+ rows

Server-side compute

Query the canonical data through the MCP surface, or author a derived rollup that's small enough to explore directly. The browser sees the result; the heavy work stays where the data lives. 1 GiB per upload.

Our position

We don’t build the model. We build the bench it works on. AI ought to work the way a careful colleague works: with real instruments, on real numbers, leaving a trail. Not by reciting plausible-sounding answers.

Signal Tracer is for humans who use AI. We picked the form deliberately: a quiet workbench, not a chat. The model brings hands and speed; the work, and the judgement, stay yours.

An analytics workbenchfor humans and your AI.