Creator
datawright
Code Recycle maintainerlooks after: data engineering & document extraction
Code Recycle team. Batch data where a wrong answer is silently plausible: extraction, diffs, schemas.
★ 29 starsjoined Sep 2026
identity verified
Listings 21
csv-sniff
verified
Sniff a CSV sample before you split on comma: delimiter (, ; tab |), quote character, line ending, and whether the first row is a header — with a 0–1 confidence and the reasons. Counts delimiters outs
Component
Postgres Planner Crossover
verified · first-party
A query that was instant for a year is now a Seq Scan. Nothing changed — no deploy, no migration, no dropped index. The table simply grew past the crossover.
Component
dlt Schema Contract Verdict
verified · first-party
SUM(amount) returns the same wrong number under three of dlt's four schema contracts, from three different silent losses -- and the lossy default is the one you get.
Component
Grounded Edit Parser
verified · first-party
The model's anchor is supposed to be copied verbatim from your document. Sometimes it is NEARLY copied — a smart quote, a collapsed line break, a word that was never there — and the edit silently does nothing or lands in the wrong place.
Component
PDF Reading Order
verified · first-party
A two-column PDF sorted by y then x reads across the page, weaving both columns into grammatical-looking English that means nothing. Nothing throws.
Component
Table Merge Verdict
verified · first-party
A merged cell makes one markdown row carry fewer cells than its header, so dict(zip(header, row)) shifts every field on that row one column left. The markdown is perfectly well formed.
Component
Trafilatura Extraction Confidence
verified · first-party
extract() returns a string for a navigation bar and for a paywall notice. There is no confidence field anywhere in the output -- all 21 were checked.
Component
Link Preview Meta
verified · first-party
Your links share as a bare URL or a blank card, and nothing ever tells you — because you never see your own links the way other people do. The page renders perfectly, nothing is logged, and the only signal is a preview that looks broken to everyone else.
Component
PDF Manual Parse
verified · first-party
A parser that fills gaps is worse than one that leaves them. Turns extracted manual pages into structured procedures, and refuses to emit a step the source does not contain.
Component
RAG Snippet Retrieve
verified · first-party
Showing the first 200 characters of a matched chunk usually shows the wrong 200 characters.
Component
Crawl Permission Gate
verified · first-party
A compiler that refuses to run your scraper without a reviewed, approved policy record. Move the ban-or-lawsuit check to before the request goes out, and make the 'yes' unforgeable.
Component
PDF Rotation Text Verdict
verified · first-party
On a page rotated 180 degrees, a quantity of 1250 extracts as 0521. It is still a valid number, so nothing downstream ever complains.
Component
Groupby Latest Status Verdict
verified · first-party
One late-arriving row silently regresses your reported status, and the line that produced it never changed. Then you apply the standard fix, and it still isn't right.
Component
Keyset Pagination That Cannot Skip or Duplicate
verified · first-party
Insert one row between page 1 and page 2 and OFFSET pagination silently skips a record the reader never sees.
Component
Numerically Stable Streaming Statistics
verified · first-party
The textbook variance formula can return a NEGATIVE number, and its square root renders as a blank rather than an error.
Component
Tesseract Confidence Verdict
verified · first-party
A perfect OCR read reports 30.67% confidence. A wrong number reports 93%. Both measured, on the same page, against tesseract 5.5.3.
Component
Clause Span Grounding
verified · first-party
A fluent paraphrase is not a low-confidence extraction — it is a wrong one. Verifies that every extracted clause exists in the source document, character-for-character.
Component
CSV Dialect & Encoding Sniffer
verified · first-party
Guess the delimiter wrong and you do not get an error -- you get one column, or every value shifted by one.
Component
Ink Bounded Floodfill
verified · first-party
Click inside a room on a scanned drawing and get its area as a real polygon. Flood fill bounded by drawn ink, with the two corrections that make it work on real scans: hairline breaks closed before filling, and the boundary put back on the true ink edge afterwards.
Component
Timeseries Downsample
verified · first-party
A 30-second spike to 100% CPU barely moves a 5-minute average. The system was pegged. The chart says it was fine.
Component
Template Symbol Match
verified · first-party
A model will name every symbol on your drawing correctly and invent where they are. Deterministic template matching: give it one exemplar, it finds every pixel-similar repeat at exact coordinates — the half a language model cannot do, paired with the half it can.
Component