Blog

A fictional inquiry gave a budget, a date, and no email. The draft sounded ready to send. The missing list caught the gap, and the reply did not.

Day 5 let a total of 9.00 pass because its quote was on the page. The new check compares the value with the amounts in its own quote and flags the mismatch. Day 6 of building a document extractor.

Intake Desk will turn a messy legal inquiry into a draft record a person can check. I did not call a model yet. I built the front door, the desk, and a place for the original message to live.

Invoice 1 passed with no issues. A quote I invented was flagged. Invoice 3 stored 9.00 while the page says 90.00, and that one still passed. Day 5 of building a document extractor.

The table now requests one page through a mock API. A slow failed request cannot overwrite a newer succeeded filter because each result has its own cache key.

The same kind of question had to be answered from the stored cards, and each answer had to name its page. One came back right, one honest and incomplete, one a correct refusal. Day 4 of building a docs assistant.

A total of 125 becomes 125.00. February 30, a negative amount, and three decimal places are rejected, and a missing currency stays null. Day 4 of building a document extractor.

Typing in the search box changes nothing until I press Search. Status, sort, and page live in the URL, so Back and refresh restore the table. Day 3 of building a dashboard for recorded application runs.

A text PDF came back with the vendor and the total. A scanned PDF came back as an empty string, even though I can read the page myself. Day 3 of building a document extractor.

At 390 pixels wide, the run table scrolled sideways and the rest of the page stayed still. I removed that rule and the whole page started sliding. Day 2 of building a dashboard for recorded application runs.

I pointed the app at a file that does not exist, and I passed the number six in quotes. The page died on the first mistake and stayed up for the second. Day 1 of a dashboard that reads recorded steps instead of guessing why a run failed.

270 docs pages became 905 searchable cards with embeddings in Supabase. The cutter is deliberately simple, and some cards are cut in the wrong place.

A missing invoice field is null, not an empty string and not zero. Day 2 of building a document extractor that keeps the value, the exact quote, and the page for every fact.

I downloaded 270 pages of official TanStack Query docs and 962 answered GitHub discussions. The docs are the library, the questions are the quiz, and many marked answers are already out of date.

One practice invoice never stated its currency, so the correct draft leaves that field null instead of guessing. Day 1 of building a document extractor that keeps the evidence for every field.

I asked an AI five questions about TanStack Query v5. Two answers used removed or renamed APIs, onSuccess and cacheTime. Day 1 of building a docs assistant that looks up the current instructions first.

How I built a production support ticket classifier using Python, rule-based routing, OpenAI API integration, and evaluation metrics. A deep dive into the boundary between custom code and external APIs.