Your photos are never training data
The most valuable asset we touch is the one we refuse to use. It is in the terms, not the FAQ, so that breaking it would be a breach rather than an update.
About us
The average phone holds
The average phone holds
eleven years of life.
Almost none of it will ever be looked at again. Not because it does not matter, but because nobody has a free weekend to sort fourteen thousand photos. We built keepsi.ai to close that gap.
Our mission
Make it effortless for anyone to turn a lifetime of photos into something they can hold.
Not a feed. Not an algorithmically resurfaced “memory” that interrupts a Tuesday. An object, on a shelf, that outlives the phone, the cloud account and the app it came from.
Digital photos are stored perfectly and remembered badly. A printed book is the opposite: fragile in theory, and the thing families actually open twenty years later. We think the future of personal photography looks less like infinite storage and more like better editing, and editing is finally something software can do well.
The Whole Year2026
14,286 photos
One book you open
The library keeps growing
A phone kept for five years holds years of life. Storage is cheap and getting cheaper, so nothing is ever deleted and nothing is ever chosen.
Almost none of it is seen again
Not because it does not matter. Because the only way back in is scrolling, and scrolling through a decade is not something anyone does twice.
A book closes the gap
Forty pages you can hand to someone gets opened more in a year than ten thousand files get opened in a decade. The bottleneck was never storage.
Why we started
Choosing is the hard part, and it is the part software can finally do.
Every tool in this category assumes you already know which photographs you want. That assumption is where people give up.
Selecting is genuinely difficult for a person, because every photograph of someone you love looks worth keeping when you are looking at it on its own. You cannot hold three thousand of them in your head at once and rank them. A model can. That asymmetry is the reason this company exists, and it only became true recently.
So we built the curation layer first and the layout layer second, which is the opposite order to everyone else. The result is that our first draft is worth reading rather than worth rearranging — and once a draft is worth reading, editing it by simply saying what you want becomes the natural interface rather than a novelty.
keepsi.ai is an early-stage company. We would rather say that plainly than dress a young product in the language of an established one.
Read the library
Eight tiers of signal from every photograph, from EXIF up to scene understanding.
Find the stories
Segment into events, group them into chapters, detect what recurs year after year.
Hand you a draft you can argue with
Change any of it by typing a sentence, at the level of the book or a single frame.
Print something that lasts
Archival stock, proper binding, and a person who checks it before it goes to press.
What we hold to
Six commitments
These are constraints on the business, not slogans. Each one costs us something.
No dark patterns
No countdown timers, no fake scarcity, no subscription hidden inside a one-time purchase. Cancelling and deleting take the same number of clicks as signing up.
You can always leave
Full export of every photo and every book, any time, in open formats. A product you cannot leave is not a product you chose.
Print responsibly
FSC-certified paper, print partners in the region we ship to, and no unrequested books. A subscription that mails you something you did not ask for is waste with a business model.
A person checks the work
Every book gets human eyes before press. Automation decides what goes on the page; a person decides whether it is good enough to send you.
Say what the AI did
Every enhancement, removal and caption is listed and reversible. You should never discover in print that something was changed without telling you.
The team
Founded by PhDs and
Founded by PhDs and
leading AI researchers.
keepsi.ai is built by people who spent their careers publishing on the exact problems this product depends on. Not adjacent problems — these ones.
Conversational AI and agents
Dialogue systems, multi-turn interaction, and the evaluation of agents that have to hold context across a long conversation.
Becomes: chat that edits a whole book without losing the threadImage selection and editing
Aesthetic and quality assessment, near-duplicate detection, subset selection under coverage constraints, and instruction-guided image editing.
Becomes: 96 frames chosen out of 3,412, and the edits you asked forPersonalization and preference modelling
Learning individual taste from sparse implicit feedback, and recommendation under cold start, where the first session is all you have.
Becomes: a second book that starts closer than the first didGraph and representation learning
Structure in relational data: roles, motifs, temporal graphs, and the embeddings that make heterogeneous entities comparable.
Becomes: the story graph that decides what belongs with whatHuman-computer interaction
How people actually work with intelligent systems: when to ask, when to act, how to show what a model did, and how to make it reversible.
Becomes: a draft you can argue with instead of acceptPrint and colour science
Colour management, soft proofing and prepress — the unglamorous craft that decides whether a screen and a page agree.
Becomes: the book on the table looking like the book on the screenResearch pedigree is worth exactly nothing on its own. What it buys here is specific: we did not have to guess which parts of this were hard. The curation problem, the cold-start problem, and the question of when a system should ask rather than assume are all problems we have worked on before, and the design decisions on this site follow from that rather than from a product meeting.
Open the drive.
Whatever is sitting in your library unsorted, this is the ten minutes that turns it into something.