Data Analysis Agent

Why it matters

Most plant engineers don’t have time to write pandas glue every time they want a trend line, an outlier check, or a forecast. The data analysis agent turns plain-language requests into executed Python, inline plots, and a written interpretation — so the engineer can ask the question they actually have instead of the one the dashboard happens to support.

Capabilities

• Three effort tiers — lightweight, standard, and deep — trading speed for richer analysis patterns.
• Two-phase deep mode — a questioning pass surfaces knowledge gaps before committing to a long analysis; the answers guide the actual run.
• Automatic multi-instrument merging with timestamp-based joins and aware NaN handling when aligning signals sampled at different rates.
• Curve comparison — overlays operating points against reference performance curves extracted by the document agent, so operators can see drift against design intent at a glance.
• Self-healing code — failed runs are retried with targeted correction prompts, so deprecated APIs and edge cases don’t bubble up to the user.
• Live streaming of intermediate text, code, output, and figures into the UI as the analysis progresses.

What makes it hold up

Generated-code agents fail the same way every time: one bad library call and the whole answer is garbage. The win here was treating execution as a conversation — errors go straight back to the model with the offending line highlighted, and the retry loop has a hard budget so failure is loud and early instead of silent and creeping.

Enterprise project. Official writeup and demo link will be added once online.