Solve the mysteries in every family's shoebox. Photo Detective reads an old photo like a historian would, dates it, places it, and then brings the family into the investigation so the story is told by the people who know it.
2. Why people can't resist it
A mystery about your own blood. "Who is the man next to Grandpa?" is the most compelling question a family can ask.
The case gets sharper. Each relative's reply visibly narrows the date range and moves the confidence bars. It plays like a detective game.
It wakes up the elders. Uncles who never text reply instantly to "Do you remember this car?". The family chat comes alive.
3. How it works for real
Scan: phone camera with perspective correction, or upload. Edge and paper analysis (deckle edge, border width, print size, back-stamp) from a vision model.
Read clues: a vision model (Gemini 2.5 Pro or GPT vision via OpenRouter) detects objects and returns boxes: vehicles, signage, clothing, hairstyles, architecture, photo format. No face identification is ever requested or stored.
Reason: a frontier LLM (Claude or GPT) turns clues into dated claims with ranges and confidence, grounded in reference tables (vehicle grilles by year, typefaces, fashion timelines, paper stocks), and combines them into a date range and place candidates.
Ask the family: the LLM writes the 3 questions most likely to narrow the case, sends them via share link or WhatsApp-style message, and parses replies (any language, voice notes via speech to text) into updated evidence.
Context: retrieval over open history sources for that place and date range to write "what was happening then".
Restore: an image restoration model (open GFPGAN-style for scratches, a colourisation model like DeOldify, or an image model edit) produces restored and colourised versions, always labelled as AI.
4. Cost per use
Vision clue read: about $0.02. Reasoning and case file: about $0.03. Each family reply parse: under $0.01. Restore plus colourise: about $0.03. A full case costs under $0.12. An optional 5 second "bring it to life" clip with a video model is $0.50 to $2.50, paid only.
Free: 3 cases a month, ask up to 5 relatives. Family archive ($5 a month): unlimited cases, restoration in full resolution, shared wall for the whole family, printed case books.
5. Growth loop
Every case invites 3 to 5 relatives with a question they can answer in one tap. Each relative sees the wall and uploads their own shoebox photo. Solved cases are shared as "case closed" cards in the family chat. Every Sunday a Case of the Week pulls one unsolved photo from the archive, and each name the family confirms grows a who's-who tree that makes next week's suggestions sharper, so the archive gets more valuable the longer the family keeps coming back.
6. Risks and guardrails
No face recognition: faces are detected only to blur or crop, never matched against any database. The app never identifies strangers.
No identifying living people: names come only from family members who add them, inside the private family space. Public sharing blurs faces of anyone not marked as deceased by the family.
Honesty: every date and place is a range with confidence, labelled "estimate". Restored and colourised images are watermarked "AI restored".
Sensitive history: photos showing violence, prisons or persecution are handled with care prompts and never auto-shared.
Privacy: photos are private to the family by default, deletable at any time, never used for training without opt-in.
7. MVP build order
Upload plus a vision model returning clue boxes and reasons as JSON, rendered on the photo.
Case file with date range and place candidates and confidence bars.
"Ask the family" share links with one-tap answers that update the case.
Restore and colourise toggle with an AI label.
Family archive wall and the "case closed" share card.