A marketing-ops reporting assistant: report emails get forwarded to one inbox, an AI pass extracts and cleans the numbers, a human approves them, and every figure lands in a version-controlled master dataset.
Every month a raw Excel export from HQ landed in my inbox: inconsistent column names, mixed formats, missing values, and enough PII that pasting it into any external tool was off the table. Cleaning it by hand took one to two hours at every monthly close, plus about fifteen minutes a day once the exports went daily.
The part that finally made
me act was quieter than the lost hours. Several of us were each running a private version
of the same cleanup, and nobody could say which cleaned copy was the latest without
asking around. The obvious shortcuts were fenced off, because the file carried personal
data that could not legally leave the building. Somewhere between the third rename of
KYC Status and the second duplicate row, I stopped treating this as
housekeeping and started treating it as a system I should build.
Connector suites solve the pulling problem well, and the scale of the problem is documented: Supermetrics' 2025 Marketing Data Report found marketers handling 230% more data than in 2020, while 56% say they don't have time to analyze it thoroughly.
But a connector needs the data to leave the building, and ours could not. Spreadsheet macros stayed inside the wall, so I tried those first; every column rename from HQ broke them, and whoever wrote the macro became its single point of failure. Pasting the raw file into a chatbot was ruled out for the same PII reason before accuracy even entered the conversation. What I actually needed was narrower than any tool on the list: strip the sensitive columns before any model sees a row, extract into a fixed schema, and keep a human between the model and the record.
The scope started as a list of things it would not do. It would not become a BI suite. No dashboard engine, no charting library.
Goals: the monthly close finishes in under five minutes; PII never reaches a model, stripped in the workflow layer before extraction; and corrections compound, so a mapping fixed once stays fixed in every later run. One inbox, one review gate, one versioned store, eight standing reports.
A Zapier flow watched the
inbox, pulled the attachment into a protected Drive folder, dropped the sensitive
columns, and renamed fields so kyc_status meant one thing. A custom GPT
with a schema doc did the extraction. The pipeline took a weekend, most of it spent
pair-writing the schema doc with the model instead of typing code. Deliberately boring,
because I had zero engineering bandwidth and a compliance team with opinions.
The plumbing held. What broke was accuracy. Published benchmarks on field-level LLM extraction put the best configurations at 71–84% agreement with human coders, and my logs matched: about one field in five needed a person. Re-checking everything would have been the old job back with extra steps, so the redesign went into the review step instead. Confidence flags per field, diffs against last month, and arithmetic that never goes through the model. The web app below, built the same AI-paired way, wraps that queue in a clickable interface.
Next.js
Gemini
Vercel
Zapier
Custom GPT
The published version of this one is deliberately framed as a demonstration rather than a product, and the landing page says so: I built it because I hate manual data entry, and complaining about a five-hour weekly tax is less useful than removing it. Three claims sit behind the four screens.
Reports arrive by forwarding an email to an inbox. Nobody opens a CSV, renames a column, or pastes a range into the wrong tab at the end of a long day.
Inconsistent column names and duplicate rows are normalised on arrival, which is the quiet failure that makes a model wrong without making it look wrong.
Master Data holds the approved snapshot under version control, so the question of which spreadsheet is the latest one has a single answer.
The monthly close that took one to two hours of hand-cleaning now finishes in under five minutes, and daily snapshots, which we never had time to produce at all, became routine.
The quieter change is the one I value more. A correction made in the review queue lands as a versioned commit, so the same fix never has to be made twice, and any month's numbers can be rebuilt from the record instead of from somebody's memory.
Next.jsThe review queue, master data view and report copilot.
ReactThe UI layer.
TypeScriptIncluding the extraction schema.
ZapierWatches the forwarding inbox and moves the attachment into the protected folder.
Gemini 1.5Field-level extraction into a fixed schema, and the plain-language questions over approved snapshots.
VercelHosting and the extraction endpoint.
Nothing here is novel and nothing here needed to be. I had zero engineering bandwidth and a compliance team with opinions, so every component is one somebody else could take over.
Nine systems shipped, compliance, reporting, intelligence, content.