Jonas ForshellSenior Product Owner
jonas@jforshell.seEmail me
Working buildUser testingCase ยท Own build

Archive Lens, a historical photo catalogue

A friend researches old photographs. The files were safe in Google Drive, but what she knew about them was scattered across names, places, dates, studios and research notes. Archive Lens connects the two in one catalogue, and every original stays where it is.

Context
Historical photograph catalogue
Role
Product Owner and builder
Stage
In user testing with its first real user
Stack
React, FastAPI, Supabase, Cloud Run, Google Vision, AI on the user's own account

The short version

Result
A hosted product in user testing with its first real user. The AI step is built but not yet measured on her archive.
What I did
Designed and built a catalogue over her Google Drive archive, where Google reads the text on each card, AI on her own account suggests names, places and dates, and a person approves them.
The call
I kept the originals in Drive and made every AI suggestion wait for a person's approval.

Easier to watch than to describe

Eight screens, no sound, thirty five seconds. Today, the Drive folders being watched, a run in flight with two failures explained, a photograph reviewed from its back with the front a click away, an album page boxed into four photographs, the catalogue, a search matching a studio mark read off a card, and AI spend against its ceiling.

Built from the product's own screenshots instead of a screen recording, so it comes out exactly the same each time the interface changes.

From local experiment to hosted product

I started with text readers and a program running on my own computer, to keep the work next to the files. The text they read was inconsistent, and the setup was far too technical for the person who would use it. So I dropped the idea of a Windows application and moved to a hosted browser workflow.

Claude and Codex wrote the code. I decided the scope, reviewed what they built and tested it.

Four decisions define the product

The boundaries
  1. 01

    Originals stay in Drive

    Archive Lens keeps a link to each file, a small preview, the text read off it and what a person has approved. It never makes a second full size copy. To read a photograph it downloads the original for a moment, makes what it needs and deletes the download. Drive keeps the only original.

  2. 02

    Listing a folder costs nothing

    You can list and browse a Drive folder before deciding what to read. Reading the text off a photograph uses part of a monthly allowance, so it only happens when someone asks for it. A run takes as many photographs as the month allows and says how many it left out.

  3. 03

    Machine output is a suggestion

    Google reads the text off each card. An AI step turns that text into names, places, dates and studios, and can describe the picture. None of it becomes catalogue data until a person accepts, corrects or ignores it, and nothing a person approved is ever overwritten by a machine.

  4. 04

    AI runs on your own account

    I have no shared AI account, so AI costs never land on me. Each archive connects its own account with an AI company and sets a monthly ceiling, and at the ceiling further calls stop. Sending the picture itself is off until someone switches it on, and then only a small copy is sent.

The workflow as it stands

Connect Drive, index a folder, process as many photographs as the month allows, review each one front and back, and export what has been approved.

Archive Lens dashboard showing what needs a person today, where the archive stands and the month's Cloud Vision allowance
The dashboard opens on what needs a person today and what's left of the month's allowance.
Archive Lens review queue showing the back of a photograph with its read lines and the suggested fields waiting for approval
Review shows one photograph, front and back. Text read from either side fills the one record, and a person decides.
Archive Lens processing view showing a run in flight with failures explained
Every photograph moves through named steps, and a failure or a retry shows as a step too.
Archive Lens AI and costs view showing the archive's own AI account and spend against the monthly cap
AI runs on the archive's own account. Spend shows against its monthly ceiling, so nobody finds out later.
Archive Lens user flow from Drive Sources through indexing, processing, the AI pass, the review queue, the catalogue and settings
The user flow from Drive connection to reviewed catalogue record.

What the AI step does

Built, not yet proven

Google's Cloud Vision reads the text on a card. It doesn't know that "S. Richards, Photographer, Wellington" is a studio, a photographer and a place, and the AI step is there to work that out. It takes that text, the filename, the folder and the archive's own vocabulary, and suggests a title, subject, photographer, studio, place and date, each with a confidence. Review treats every suggestion like a line of text read off the card, so a person can accept, correct or ignore it.

The text step sends no image and is the cheapest call in the system. Describing the picture is a separate step that sends a 384 pixel copy, and it stays off until the workspace switches it on. A cheap model answers first and a stronger one is only asked when the first is unsure. Before a large run starts it shows the estimated total, and once spend reaches the monthly ceiling further calls are blocked.

All of this is built. It hasn't run over the real archive yet, so nobody knows how often it names the right studio or invents a date. Measuring that starts in the lab below.

Where the AI gets tested

September 2026

The AI lab tests the rules the model follows before they reach anyone's archive. A case is a public-domain card with its text locked and the right answers approved by hand. A run sends those cases to one model under one version of the rules, and each field it fills gets marked right or wrong.

So far that's five cards and eight runs, across four models and three wordings of the rules, for 14 cents in all. Five cards is far too few to trust a score, and the screens say so. The model answers below are real. Only five fields have been marked by hand, so a matching rule marked the rest for these screenshots.

Archive Lens AI lab experiments screen with the best score for each rules version and a cost against accuracy chart
Each version of the rules gets its best score, and the panel below weighs a dearer model against what it adds.

What the research added

August 2026

After three months with one real user I couldn't tell which of her habits came from the hobby and which were her own. So I read the guides people follow to identify old photographs, the forums where they get stuck and the complaint threads about the tools they leave. Six changes came out of that reading, and all six are built.

Search forgives spelling, because names on the back of a card are rarely spelt the way the family spells them now. Filters show which records still lack a place, a date or a subject. Review can be narrowed to one folder, since every guide says to work one box at a time. Shared fields can be set across many records at once. And every photograph gets a fingerprint of how it looks, so review says when one looks like a print already catalogued.

The biggest was the back of the photograph. Every guide treats the front and the back as separate evidence, the picture on one side and the studio stamp, tax stamp and handwriting on the other. Until then the app had no idea a back existed. Now a back is its own file with its own reading, paired to its front, and review shows both sides as one record.

Next I want to watch her review sessions and shape review, search and export around what she does in them.

Decision log

Newest first

The mistakes in this log are about building the wrong thing, or the right thing at the wrong time. I built a Windows app the user could never have run, and a test scoring screen nobody opened. Half price bulk processing went in a month before there was any volume to discount, and the research I did in August belonged in May.

Each entry is a release that carried a product call, tagged with how it turned out. The "got wrong" lines are the most useful part.

  1. Right thing

    Showed the lab with real answers and said how they were marked

    The AI lab screens on this page show the real answers from the September runs, copied out of the live lab. I've marked only five fields by hand so far, so a simple matching rule marked the rest for the screenshots, and the section says so. Those scores show how the screens work and say nothing reliable about the models.

    Got wrong

    The chart that says whether new rules did better had never counted a run on a draft. Since September it had called the first version the best and shown the two newer ones as unscored. I found it while taking the screenshots.

  2. Right thing

    Asked what each object is before building fields for it

    A redesign of the lab arrived as finished screens, and it left out that randomness setting. The setting was built first, since the first evening showed nothing could be measured without it. Anything the model works out for itself, like a country from a town, is always held for a person to check. Only staff can make new rules live for every archive.

    The first cards drawn from a real archive included a newspaper clipping. Every field was built for studio portraits, and on a clipping there's no studio, town or street, so the model is pushed to make them up. I added one question, what kind of object this is, and left the rest alone. Fields for each kind come after the archive has shown which kinds it holds and how many.

    Got wrong

    The guide and the Fields screen said fields with a good enough score are saved without a person. Nothing is saved that way yet, so both now say "would be saved".

  3. Right thing

    Kept the cheapest model and left automatic saving off

    Four models got the same rules and the same five cards. Claude Sonnet got three to five more fields right out of forty, at nearly thirteen times the price. Claude Haiku and GPT-5.6 cost four to six times as much, and their scores moved no further than the cheap model's own scores move between two identical runs. So it stays on GPT-4.1 mini, which also keeps a free allowance possible if other people ever use this.

    Each answer comes with the model's own confidence. On seven fields it was right every time it said high, 154 times out of 154, and titles were wrong all twelve times it was sure of one. Saving those seven fields without a person still waits. The 154 answers come from the same five cards, and five cards can't rule out a model that's right only six times in ten.

    Got wrong

    Three of Sonnet's five misses were arguable answers on my side. It said England where I'd written United Kingdom.

  4. Right thing

    The first real test found mistakes in my own answers

    The AI lab compares the model's answers for a card with answers I've approved by hand. The first evening of real use went through eleven cards, and the first problems were in my answers. A photography firm's name sat in the person field on six of them, so a model that correctly left the person blank would have been marked wrong. My answers also disagreed about whether a town is enough to name the country, eight cards saying yes and three saying no for the same reason. Country now follows one rule on every card.

    Then I ran the exact same test twice. The two runs disagreed on seven of 43 fields, as many as the biggest change to the rules I'd tried that evening. Nothing measured that night could be told apart from chance, so the next build began with a setting that turns down the model's randomness.

    Got wrong

    I nearly wrote up a change to the rules as an improvement. Running the same thing again cost half a cent and showed it was chance.

  5. Wrong time

    The database lock stopped being optional

    Since August the database itself decides which archive a request may touch. That was rolled out behind a switch that was off by default, so it could go to the test site first, be watched for a wrong rule and then go live. That order was right. Leaving the switch optional afterwards wasn't. A review of the code found that nothing stopped the live site from running with the lock off, and nothing recorded which sites had it on.

    So it's no longer a choice anywhere people can reach. The online app now refuses to start unless the database lock is on. Both sites already had it on, which took a manual look to find out. The same review fixed a background job that had failed every fingerprint for want of one setting, and the browser tests now run the whole system together so a miss like that can't pass twice.

    Got wrong

    The rollout switch was left as a permanent option. A default that says off proves nothing about what's live, and I had no way of knowing without going to look.

  6. Wrong thing

    The AI had four screens and no front door

    Turning it on, paying for it, judging it and tuning it were spread over four screens, with two buttons that started a run and two tables scoring the same thing. Nothing anywhere said where the results appear. Two different people were sharing those screens. Someone enriching their own archive needs four steps in a line. Whoever decides what the archive sounds like needs a workbench, and that's a different job.

    So they were split. Enriching became a numbered step between processing and review, ending with a sentence saying the suggestions wait in review. The rules the model is given stopped being part of the code and became something a person can edit, try against prepared cases, and publish when it's better. Only a named account sees that, and the rest of the archive never learns it's there.

    Got wrong

    The second run button had never worked. It sent a request the server refused, and because the first button worked nobody noticed, including the test written to cover it. Removing the duplicate was the fix. The same week showed the automatic check over the whole codebase had stopped running on the machine it was meant to protect, and it was hiding a real fault.

  7. Right thing

    The demo learned the back of the card

    The demo script now pairs each back to its front the way review does. The case study says what the AI step does, and the flow diagrams show the AI after the text is read, never in place of it. This log started the next day.

  8. Right thing

    Five rounds on one sentence

    The duplicate scan said every photograph had already been fingerprinted while thousands had none, three times, each time for a different reason. Each fix had added a skip without adding the number that made it visible. The answer was six counts that between them cover every photograph without a fingerprint, so the all-clear can only appear when it's true.

    Got wrong

    Twice more on the way. A background job timing out looked like an unreadable photograph and would have retired it for good, and "retry the failed ones" was retrying everything.

  9. Right thing

    Six features, one wave each

    Search that forgives spelling. Filters for records still missing a place, a date or a subject, because what's left to do matters more than what's done. One folder at a time in review. Bulk edit that goes through the same checks as editing one record. A fingerprint of how every photograph looks. And the back of the photograph as its own file, paired to the front, with review showing both sides of one record.

    Got wrong

    Making more fields searchable also changed what the archive showed. On the live database seven fields stopped being searchable without any warning, while the test database kept finding them. The fix keeps what's shown and what's searched apart, and tests every field on the same kind of database the live site uses.

  10. Wrong time

    Found out who the user really is

    Three months in, the feedback from one person was thin. Not because the product was fine, but because she's one person with a life. I couldn't tell her habits from the hobby's. So I read the guides, the forums and the complaint threads of the tools people leave, and treated that as the second user. Six features came out of it. The plan started by tidying the code, because the rule for what waits in review was written in five places and the filters in three.

    Got wrong

    I waited for feedback to arrive instead of going to get it. The research should have been done in May.

  11. Right thing

    The database keeps archives apart, and one visual style

    Keeping each archive private moved into the database itself. Automatic tests prove it on every change, signing in the real way against the real rules. Every change is now checked for leaked passwords and keys. The old stylesheet was deleted, and a third clean-up finished the pieces the redesign left behind.

  12. Right thing

    The product photographs itself

    A script fills a throwaway archive with public-domain cartes de visite, steps through eight screens and saves a picture of each. A silent video is built from those pictures. A screen recording drifts and goes stale. This one is rebuilt exactly the same way whenever the interface changes, and it lets people see the product without needing the first user's time.

  13. Right thing

    Nine things raised from real photographs

    The first user's cards were sideways, held scanner marks read as letters, and one album page carried four people's captions in one record. So, turn a photograph, mark a line of read text as not on the card, read it again for real, fix a sideways preview without paying for a read, and split a page into records of its own.

    Got wrong

    Reading a turned photograph again would have handed back the same sideways text for free, because the app saw the same file in Drive and reused the old reading. Turning a photograph doesn't change the file in Drive. And nothing in the background was set up to pick up those two new kinds of job, so they were queued and never done. I found both by trying it.

  14. Right thing

    Redesigned from a written brief

    A brief for AI design tools set the hard rules first. Originals never leave Drive, machines suggest and a person approves, and the page never jumps when a temporary control appears. The interface was rebuilt to the result in one go. Notes became separate lines the same week, each remembering which reading it came from.

  15. Right thing

    A user is a user

    I removed the admin role instead of securing it. Each person owns a private workspace, so there was nothing for an admin to be admin of. The same day added a limit on how many photographs one click can queue, and the standard browser security settings on the web app.

  16. Right thing

    A fourth AI service, kept to text only

    OpenRouter was added so free AI models could be tried. It handles text only, with no bulk option, and nothing in the code can send it an image. Its list of models is fetched live, because the free ones change too often to type in by hand. It can't promise that nobody keeps the data sent to it, so its settings card says so in plain words instead of forcing every request through the few services that can.

  17. Wrong thing

    Deleted what nobody could reach

    A clean-up removed a screen that scored the AI against test photographs, and four other pieces nothing used any more. Each removal is written down with the reason, and the bulk processing code that stays is marked as kept by choice, so the next clean-up doesn't mistake it for something forgotten.

    Got wrong

    The test scoring screen was a whole feature built to answer a question the accuracy report already answered. Nobody opened it. The test photographs came back with the August redesign, as a list someone vouched for, and their screen says plainly that the comparison doesn't exist yet instead of showing a button that does nothing.

  18. Right thing

    Every failure gets a reference

    Deleting a record now clears everything attached to it, thumbnails included. Every error shows a reference number, crashes in the browser report themselves, and Settings lists what was recorded. Built because "it did not work" from a non-developer is a dead end without something to quote.

  19. Wrong time

    AI on the user's own account, built in a day, parked in two

    An AI step turns the text read off a card into names, places and dates. I have no shared AI account, so the cost can only land on the archive's own account. Text came first, because that's where those fields come from. Describing the picture sends a 384 pixel copy and stays off until someone switches it on. A run shows an estimate first, stops at a monthly ceiling and reports how accurate it was afterwards.

    Half price bulk processing from all three AI companies was built the same day. Two days later I decided not to switch it on, and wrote down what would make me look at it again. Half price is real money at volume, and this archive isn't volume.

    None of it has been run over the real archive yet. The accuracy report and a set of test photographs exist to measure it, and that measurement is still ahead.

    Got wrong

    I built the bulk discount before there was anything to discount, two days of work the plan for one user never needed. And a month later the AI step is still unproven on real cards, which is the only proof that counts.

  20. Right thing

    Made it safe before it could be shared

    A security review came before a second person touched it. Each person got a private space, and every photograph went back to its real owner. Connecting Google Drive now works only once, in the browser that started it. The database's internal tables were locked to everyone by default, and the old prototype that ran on my own computer was deleted instead of being kept as a second way in. The live site wasn't touched until a test copy had proved all of it.

  21. Wrong time

    Put it down and built something else

    The MVP was in the first user's hands. It was her idea, and she's hard to get time from, so waiting for feedback would have meant weeks of nothing. I went and built Resume Builder instead and let Archive Lens sit until there was something to react to. Eight weeks went by without a change to the code, and I chose that.

  22. Right thing

    A dense workbench, not a brochure

    A design spec from Stitch set three panels side by side, search across all the text, and a back button that remembers your filters and the record you had open. The rule written down then still decides how the app looks. It should feel like a work tool, stay plain and always say what happened and what to do next.

  23. Right thing

    Processing became a cost decision

    Google's servers download an original for a moment, make a preview and a copy to read the text from, send that to Google's Cloud Vision and delete the download. The last text reader running on my own computer was removed on the 10th.

    Reading got a monthly cap of a thousand photographs, which can be changed, and the option to bill it to the user's own Google account. Listing a folder became a separate step from reading it, so nothing uses up the allowance just because it's in a folder.

  24. Wrong thing

    Started local, left the same day

    I compared three text readers on real cards, with a program running on my own computer and a review screen beside it. The readings were inconsistent and the setup needed a developer. The app was running online before the day was out.

    The call that outlived everything else was made here too. Originals stay in Google Drive, and the app only ever keeps links to them and smaller copies made from them.