Nothing to hand? Load the (boringly regular — the right answer is regex and zero model calls), the (mostly regular with real damage — the right answer is a hybrid pipeline), or the (judgement calls — the right answer is LLM-first with cost containment). Saved example runs replay for free.
Paste a sample — the profile is free
No upload, no AI: the profile reads your sample in the browser and measures what it mechanically can. How much of the text follows one line shape (the consistency number the whole decision hangs on), whether a delimiter with a stable column count is hiding in there, which record markers repeat, and what noise — page numbers, ruling lines, encoding junk, OCR hyphen-wraps — will poison patterns and prompts alike. This costs nothing and happens while you type.
The AI writes the plan — this is the metered part
A senior data engineer's pass over the same sample: an honest verdict — regex-first, hybrid, or LLM-first — with the evidence named, the actual patterns with capture groups and expected match counts, ordered cleaner rules, the mechanical confidence checks that decide which records escalate, an LLM stage sized to the cheapest tier that works, the pipeline, and the edge cases. Every profile flag gets confirmed or explicitly set aside. Pricing is honest: a worst-case amount is reserved before the run and only what the run actually uses is charged — the meter next to the button shows both.
Verify in place, export, build
Press “Test all on my sample” and every proposed pattern is compiled and executed against your paste, right here — the plan's expected counts against the real ones, disagreements surfaced, nothing taken on faith. Then take it home: the full plan as Markdown or JSON, the patterns as patterns.json, and a starter parser in Python or JavaScript with the cleaner rules and patterns already wired. Plan history stays on this device with restore, so you can re-plan after fixing the upstream export and compare.
Derived from the @affaan-m/regex-vs-llm-structured-text skill.