Agentic AI
PastPresent
Turning scattered memories from dozens of contributors into one printed keepsake — where the automation has to stay invisible for the book to feel handmade.
5Occasion Types Supported
At a glance
The engagement
- Client
- PastPresent
- Industry
- Agentic AI
- Engagement
- Product build
- Stack
- AWS Amplify, RAG, AI Agents
- Team
- Three senior engineers
- Status
- Shipped, in production
- 5Occasion Types Supported
- RAGGrounded in Real Memories
- ZeroApps for Contributors
- Pre-PrintAutomated Moderation
Before
The situation when they came to us
PastPresent collects stories and photos from everyone who knows a person — for a memorial, birthday, wedding, graduation, or anniversary — and turns them into a hardcover tribute book. The product worked, but it did not scale: every additional contributor added manual editorial effort. Writing quality varied wildly between contributors, people got missed, layouts suffered, and inappropriate submissions could reach print. ENGISOLS led the technical side as CTO and built the AI backbone end to end.
Editorial effort scaled linearly with contributors — every extra person meant more manual writing, chasing, and layout work.
Contributions arrive in wildly different registers, from a two-line message to a long reminiscence, and the book has to read as one coherent volume.
Levelling the writing risks flattening it — the whole point is that each entry still sounds like the person who wrote it.
Generic AI writing invents details. For a memorial book, a plausible-but-false memory is not a glitch, it is a serious failure.
Submissions are unmoderated and the output is physically printed, so inappropriate content has to be caught before it becomes permanent.
Photos arrive unsorted across decades and have to be arranged into something that reads as designed rather than dumped.
What we built
An agentic AI pipeline that turns crowdsourced memories and photos into a printed tribute book — grounded by RAG, moderated before print, invisible by design.
Technical decisions
What we chose, and what we gave up for it
Grounding — RAG Over Each Person's Memories
Retrieval-augmented generation scoped to the individual honouree, so every generated line is anchored to details contributors actually submitted. The model works from a real corpus of that person's life rather than generating plausible filler — the difference between a book that moves a family and one that quietly gets things wrong.
Agentic Pipeline
An agent pipeline carries a submission from raw contribution to print-ready page: interpreting what was sent, levelling the prose, placing it in the book, and handing off to layout. Each stage is a distinct step rather than one monolithic prompt, so quality problems can be isolated and fixed.
Voice-Preserving Editing
The editing pass raises the floor without flattening the ceiling. Grammar, structure, and pacing get levelled so a two-line note sits comfortably beside a long reminiscence, while the phrasing and character of each contributor survive the edit.
Photo Curation
Automatic photo arrangement across the book, curating by theme and time period so images land near the stories they belong to instead of being dropped in submission order.
Moderation Before Print
A moderation stage screens submissions before anything reaches print. Because the artefact is physical and often commemorative, this runs as a hard gate rather than a post-hoc flag.
Infrastructure — AWS Amplify
The platform runs on AWS Amplify, covering the contributor-facing submission flow, the organiser dashboard, and the pipeline behind them.
Stack
- AWS Amplify
- RAG
- AI Agents
Results
What changed
- Editorial effort decoupled from contributor count — adding people no longer adds proportional manual work.
- Stories stay factually grounded in what contributors actually wrote, not model invention.
- Contributors need no account, app, or payment to take part — an invitation link is enough.
- Inappropriate content is caught before print rather than discovered in a finished book.
- The automation is deliberately invisible: the finished book reads as though it was assembled by hand.
