Lore — The Stories That Made You, Saved Forever
Inspiration
My grandfather fought in a war, survived a famine, and raised seven children in a village with no electricity. He spoke three languages, cooked food I have never been able to replicate, and had opinions about everything from cricket to politics to the right way to tie a turban.
I never recorded any of it. He passed away before I thought to ask.
This is not a unique story. It happens in every family, in every country, in every generation. The people who lived the most extraordinary ordinary lives — the grandmothers, the grandfathers, the uncles and aunts who shaped who we are — leave without their stories being captured. Not because nobody cared. Because nobody knew how to start the conversation.
The problem is not indifference. The problem is distance, language, awkwardness, and the quiet assumption that there will always be more time.
We built Lore because there is not always more time.
What We Built
Lore is an AI-powered oral history platform that interviews elderly family members in their own language through gentle voice and text conversations, then automatically transforms those conversations into beautifully illustrated digital memoir books that families keep forever.
A family member creates a profile for their elder — their name, age, language, relationship, hometown, and interview style. Lore generates a private interview link. The elder opens the link on any device with no app download and no login required. The AI greets them warmly in their own language and begins a gentle conversation — one thoughtful question at a time.
When the session ends, Lore extracts the stories from the transcript, writes them as first-person memoir prose in the elder's own voice, generates a culturally-matched watercolor illustration for each story, and organizes everything into a beautifully designed digital memoir book with chapters, drop caps, poetic titles, and a downloadable PDF.
Families can read the memoir in any of 12 languages, listen to stories read aloud, favourite specific memories, share a public memoir link, and receive AI-generated suggestions for what to explore in the next conversation.
How We Built It
Lore is built entirely inside MeDo — Baidu's AI-powered no-code platform. No traditional code was written. No external APIs. No paid services. Everything runs through MeDo's skill ecosystem.
We orchestrated 8 MeDo skills working together as a unified system:
The 8 MeDo Skills
| Skill | How We Used It |
|---|---|
| LLM (×2) | AI Interviewer (conversation) + Story Extractor (memoir writing) |
| Speech-to-Text | Elder voice input during interviews |
| Text-to-Speech | AI voice responses + "Listen to story" memoir playback |
| Image Generation Lite | Culturally-matched watercolor illustrations + memoir covers |
| Google Text Translation | Multilingual UI, translated memoir view, TTS pipeline |
| Downloadable memoir book generation | |
| Login | Family member authentication |
| Word | .docx export option |
The AI Interviewer
The LLM skill powers a carefully engineered interviewer persona. The system prompt is dynamically constructed from the elder's profile — their name, age, relationship, hometown, language, interview style preference, and any topics the family requested be explored. The AI responds exclusively in the elder's language, uses culturally appropriate honorifics, and asks exactly one question at a time.
We designed three interview styles — Gentle (softer tone, simpler questions), Curious (deeper follow-ups and exploration), and Reflective (questions about meaning and lessons learned).
The Story Extraction Pipeline
A second, entirely separate LLM invocation transforms raw conversation transcripts into first-person memoir prose. It identifies 1 to 5 distinct stories from each conversation, writes them in the elder's voice, assigns a poetic book-chapter title to each, classifies them into one of 6 life chapters (Childhood, Family & Marriage, Work & Purpose, Hardships, Lessons & Wisdom, Surprises), and outputs strict JSON for the app to parse and display.
The Cultural Image Generation System
One of our most ambitious technical decisions was building a cultural context injection system for image generation. Every elder profile maps to a set of cultural visual descriptors based on their language — covering people, architecture, landscape, clothing, and visual style — which are dynamically injected into every image generation prompt.
A Russian elder gets birch forests, snow, wooden izba houses, and traditional Russian clothing rendered in a painterly watercolor style influenced by Russian artistic tradition. A Hindi-speaking elder gets Indian courtyards, mango trees, monsoon fields, and culturally appropriate clothing in warm golden-hour Indian memoir style. A Japanese elder gets cherry blossom trees, tatami rooms, and woodblock-print-influenced watercolor aesthetics.
Gender is detected automatically from the relationship field so a grandfather always receives an elderly man illustration and a grandmother always receives an elderly woman illustration.
The Translation Architecture
Google Text Translation skill is used in four distinct ways:
Multilingual interview UI — all labels, buttons, error messages, and end-of-interview screens are pre-translated on session load and cached for the entire session
Language toggle — elders can switch the interview experience between their native language and English at any point, with all transcript text and UI labels toggling simultaneously
Memoir translation — family members can read any story in any of 12 languages with translations cached in the story record after first generation
TTS pipeline for Hindi — since Hindi TTS voice models produce unreliable output, Hindi AI responses are first translated to English then spoken via English TTS, ensuring clear intelligible audio
The Memoir Design System
The memoir view was designed to feel like a leather-bound book in a sunlit room — warm bone background (#F8F4ED), sienna accents (#A0522D), soft gold highlights (#D4A574), Cormorant Garamond serif headlines, drop caps on every story opening, chapter dividers with ornamental elements, and illustrations displayed as full-width watercolors above each story.
Challenges We Faced
1. Making TTS Work Across Scripts
Text-to-Speech for Indic and non-Latin scripts was our most persistent technical challenge. Languages like Hindi, rendered in Devanagari script, produced completely unintelligible audio when passed directly to the TTS skill because the underlying voice model had no native support for that script.
Our solution was to build a translation-intermediary TTS pipeline — the AI response is first translated to English using the Google Translation skill, then the English text is spoken via TTS with an English voice model. The original language text remains visible in the transcript while the audio plays in clear, intelligible English. A note informs the user that audio is playing in English translation.
2. Preventing Interface Deadlock
Early in development the interview screen would enter a permanent frozen state if an LLM or STT API call hung. The user had no way to escape without refreshing the page, which lost their entire conversation.
We solved this by completely decoupling TTS from the conversation state machine, implementing hard timeouts on every API call (15 seconds for LLM, 10 seconds for STT, 8 seconds for translation), building a cancel button that appears the moment any spinner appears, and enforcing a guaranteed reset to IDLE state on every page load regardless of prior session state.
The text input is now unconditionally active in every possible application state. There is no scenario in which the elder cannot type a response.
3. Cultural Accuracy at Scale
Building image generation prompts that produce genuinely culturally accurate illustrations across 12 languages and 12 distinct cultural contexts — and ensuring gender accuracy from relationship field parsing — required building a systematic cultural descriptor mapping rather than relying on the image model's defaults. The difference between a generic illustration and one where a Russian grandmother is shown sitting by a frost-covered window with birch trees and an Orthodox church visible outside is the difference between a product that feels real and one that feels like a demo.
4. Designing for the Least Tech-Savvy User
The elder receiving the interview link may be 80 years old, unfamiliar with smartphones, and reading in a non-Latin script. Every design decision was made with this user in mind — no login required to start an interview, font sizes adjustable during the session, all UI labels pre-translated to the elder's language before the page renders, clear single-purpose buttons with both icons and text labels, and a conversation layout that looks like book passages rather than a chat application.
5. HTML Entity Decoding in Translated Content
When translated strings from the Google Translation skill were
inserted into the UI, HTML special characters were rendering
as raw entities — apostrophes appeared as ' and quotation
marks as ". This required building a universal decode
pass applied to every translation output before display across
the entire application.
What We Learned
Building Lore taught us that the hardest problems in AI products are not the AI problems — they are the human problems.
Getting the LLM to respond in Marathi was straightforward. Getting an 80-year-old grandmother in Mumbai to feel comfortable enough to share a memory she has never spoken about before — that required thinking deeply about tone, pacing, honorifics, cultural context, and the subtle difference between an interview and a conversation.
We learned that multilingual is not just a technical feature. It is a statement of respect. When an elder opens a link and sees the AI greet them in their own language, using the right honorific, asking about the home they grew up in — that is the moment Lore stops being a product and becomes something that actually matters.
We also learned that MeDo's skill composition model is genuinely powerful for building products with real emotional depth. Orchestrating 8 skills — LLM, STT, TTS, Image Generation, Translation, PDF, Login, and Word Export — into a unified user experience that an elder can navigate without any technical knowledge is something we could not have built this quickly, or this accessibly, any other way.
The Bigger Picture
There are approximately 1 billion people over the age of 60 alive today. By 2050 that number will exceed 2 billion. The overwhelming majority of them have stories that have never been recorded.
If even 1% of families used a tool like Lore, that is 10 million life stories preserved — stories of wars survived, countries left behind, children raised in impossible circumstances, loves found and lost, recipes never written down, lessons learned the hard way.
Lore is not an app. It is infrastructure for human memory.
Built with MeDo. No traditional code. 8 AI skills. 12 languages. One mission — the stories that made you, saved forever.
Built With
- medo-(baidu-ai-no-code-platform)
- medo-google-text-translation-skill
- medo-image-generation-lite-skill
- medo-llm-skill
- medo-login-skill
- medo-pdf-skill
- medo-speech-to-text-skill
- medo-text-to-speech-skill
- medo-word-export-skill
Log in or sign up for Devpost to join the conversation.