Parview was an enterprise document-archiving tool we built in the late 2000s. It's long retired — document management is a mature, well-served category now and we don't work in it any more. I'm leaving this page up because the product is gone but the lessons aren't, and three of them have proven true on every project I've run since, including the AI work that pays the bills today.
How it started
By the mid-2000s a lot of organisations had worked out that their filing rooms were a liability with a deadline on it. Decades of paper — student transcripts, property deeds, court filings, medical records — sitting in basements being eaten by damp, mice and staff turnover. The people who knew where anything was kept retiring, and the people replacing them couldn't find a thing.
We did custom digitisation projects for a few universities and government offices. By the third or fourth one it was obvious the software layer barely changed between clients: scanner integration, OCR, encrypted storage, full-text search, access control. So we packaged it as Parview. Standard consultancy move — productise the thing you keep rebuilding.
Lesson one: scanning is the easy part
On the first project we scoped scanning as roughly 80% of the effort. It was closer to 20%. The actual work — the part that ate the timeline — was metadata. Tagging each document with the context that makes it findable later. You can OCR a handwritten student record from 1958 perfectly and it's still lost forever if nobody marks it as student records, 1958, mechanical engineering. The bits are not the point. The findability is the point, and findability is human labour. I have watched teams make the identical miscalculation with AI projects ever since: the model is the easy 20%, and the data work nobody wanted to scope is the other 80%.
Lesson two: build around the librarian, not over them
Every filing room has a librarian — the person who can locate any document in minutes using a mental model they cannot fully explain. The tempting move is to replace that person with the software. It's the wrong move, and it's expensive to learn the hard way. The correct move is to build the software around their mental model, with them in the room. The deployments where we sat with the librarian and encoded how they actually thought were used and loved. The ones where we shipped over the librarian's head were shelved inside a year. Same code. Different outcome, entirely down to whether we treated the expert as a problem or as the spec.
Lesson three: permissions are the whole game, from day one
A digital archive gets used by far more people than a paper one, because it's easy. That's the entire value proposition. It's also the trap. In the paper world, sensitive records were protected by the physical friction of having to walk to the filing room and ask. Digitise it and that friction vanishes — and a single mistake in access control now exposes things that were effectively private by accident. We learned to build role-based permissions on day one, down to the individual document, never as a phase-two bolt-on. Retrofitting security onto a system that was designed without it is one of the most reliably painful exercises in this industry. Don't.
What it actually was
A web-based archive. You logged in, searched, got your documents back. Under the hood, the parts that mattered:
- Encrypted storage. AES-256 on everything at rest, decrypted only with the user's passphrase.
- Encrypted transit. SSL, back when configuring it still counted as a day's work.
- Geographically redundant backups. Three continents. Looked excessive on the invoice. Saved us at least once.
- Role-based access control, down to the document level. See lesson three.
- Real search. Full-text over the OCR'd content plus every captured metadata field, ranged queries on dates and numbers, even colour queries for image-heavy archives.
The day it landed
The moment I remember from that era was watching a professor at an Indian university the day we switched Parview on for their records room. She typed in a student's name and had a full set of transcripts, letters and examiner reports back in two seconds. Then she sat back and said there was a question a student had asked her six years earlier — whether the student's late grandfather had once lectured there. She typed his name. Up came a 1963 appointment letter. She started to cry, in the way that tells you someone's been carrying an unanswered question for a long time.
We hadn't built it for that. But it's the clearest reminder I've had that infrastructure work touches people in ways you can't design for in advance — if you build it carefully enough to let them in.
Why this still matters to me
Parview is retired and we don't build archives any more. But the three lessons travel intact onto everything we do now. The model is the easy part; the data and the metadata are the work. Build with the expert who already knows the domain, not over them. Permissions and privacy from day one, never bolted on. None of that depends on whether the capability underneath is a full-text index or a large language model. The technology keeps changing. The mistakes stay remarkably consistent.
