Skip to content
AI & Automation

Dead Archive-to-Training Data

Decades of design and engineering files sat on an on-premise archive that nobody could search, because the filenames meant nothing and the people who understood them were busy. The knowledge existed, spread across whoever happened to remember a project. We built a game that pays people to write it down, and exports the result as a dataset an AI can read.

Snapshot

Subject
An on-premise design and engineering file archive
Engagement
Catalyst — scoped build
Timeline
~3 months
Scope
File serving, scoring, referral chains, consensus gating, dataset export
Coverage
48 file types across three difficulty tiers

Stack

  • Node.js
  • Express
  • Markdown export
01

The problem

An archive is only an asset if you can find things in it. This one held decades of drawings, models, and documents whose filenames carried no useful meaning, so retrieval depended on asking the one person who might remember. Nobody was going to sit down and label thousands of files as an unpaid side task.

  • Filenames carried no reliable description of contents
  • Recognition was distributed — no single person could describe the whole archive
  • Specialist formats needed someone who owned the software to identify them at all
  • Any bulk labelling effort competed with real work and would lose
02

What we built

A server that hands out one file at a time and asks the person in front of it what it is, with inline previews for images, PDFs, and text so the easy ones take seconds. A 10-word minimum keeps a description useful.

Scoring tuned to difficulty: 10 points for images and text, 20 for documents, 40 for the 24 specialist formats — Revit, SketchUp, DWG, IFC, STEP, SolidWorks and the like — where identifying the file needs someone who works in that tool.

A referral mechanism for files you do not recognise: pass it to a colleague, and the multiplier rises with the chain, 1.0 to 1.5 to 2.0 capped. Whoever finally cracks it earns base × multiplier, and everyone who referred it collects a finder bonus, so passing a file on is rewarded rather than penalised.

A consensus gate holds points pending until two people agree on a description; disagreements route to an admin conflict queue instead of paying out. Locked descriptions export as Markdown — the training dataset for an AI over the archive, which was the actual objective.

03

What stuck

  • Describing archive files became a scored activity rather than an unpaid chore
  • Consensus and a conflict queue keep quality from degrading as coverage grows
  • Every scoring value lives in one config file, so the economy is tunable without a deploy
04

Impact

  • Archive knowledge moved out of individual memory into a checked, exportable record
  • Specialist formats got identified by the people who own the software, via referral
  • PLACEHOLDER — add a measured figure here (e.g. share of the archive described, or files per week)