Legado | Devpost

Inspiration

Four people in our families put their ambitions on hold for decades:

A grandmother in India who wanted to be a doctor and was married after her father died.
A grandfather who spent about 40 years as a chemical engineer in India's nuclear program and wants to keep building and to write a book about purpose.
A father from Ghana who paused building a lab to pay for his children's college.
A mother whose retirement business idea stalled when the market proved hard to enter.

They are educated and use technology, but AI tools feel like a black box to them. Courses lead nowhere, and eventually they ask why bother at their age.

At the same time, young adults are losing entry-level paths. Stanford's Digital Economy Lab reports that employment for 22–25-year-olds in highly AI-exposed occupations is about 19% below expected levels. Harvard's Making Caring Common found that 58% of young adults reported little or no purpose or meaning in the prior month.

What it does

Legado runs six-week meaning externships. An older adult leads a team of 3–4 young people on a real project: a book, a course, a venture plan or a cultural archive.

Memory cards: the older adult records what they know by voice. Each recording is stored verbatim, with a transcript, topic tags and a visibility setting.
Three tracks: Teach (author or teacher), Learn (late learner with a study plan and mentor) and Build (founder with a plan the team tests).
AI that suggests and never authors: the older adult picks a drafting level (Keep my words, Smooth it lightly or Help me rewrite it) and approves everything before it leaves the circle.
Young collaborators take defined roles and finish with a portfolio piece and a reference.

How we built it

Our demo is a clickable prototype with hardcoded data: our families' stories, scripted AI responses and no backend. It walks through the full flow for all four personas, including voice intake, card review, matching, the weekly session companion and staged publishing. The rest of this section describes how we would build the production version.

Architecture. A TypeScript modular monolith (Next.js and tRPC) on Supabase Postgres, with an Inngest job queue for all voice and model work. Five modules (Cards, Circles, Drafts, Publishing, Identity) each own their tables behind typed interfaces, so any one can move into its own service without a schema change. Every table carries an org_id, so a senior center or an employer retiree program can run cohorts as a tenant, isolated by Postgres row-level security.

Voice to memory cards. Leads record on the web or by phone through Twilio, so a landline is enough. Each recording is hashed (SHA-256) and stored write-once in Cloudflare R2. Deepgram returns diarized, word-timestamped transcripts that are never edited afterward. Claude segments each transcript into cards, and a validator rejects any card whose text is not an exact concatenation of the original transcript segments. Every step is an idempotent job keyed on the audio hash.

Drafting guardrails. Model output is stored only as suggestions, and nothing enters a draft until a person accepts it. The drafting levels are enforced in code:

Keep my words allows no content-word substitutions.
Smooth it lightly rejects any new name, number or claim.
Help me rewrite it requires every sentence to cite a source card and pass a grounding check.

A voice profile built from the lead's own recordings flags passages that drift from how they speak.

Approval and provenance. The lead's approval is bound to a hash of the exact text, so any later edit sends it back for review. Each published page carries a manifest linking every paragraph to its source cards and recordings. Health treatment claims are blocked before publishing.

Scaling. Per-organization concurrency limits keep one sponsor's batch jobs from slowing live sessions. Jobs that aren't time-sensitive run through Claude's Message Batches API with prompt caching. The transcript and event tables are partitioned by month. We expect coordinator time to run out before compute does, so matching shortlists, weekly check-ins and the checkpoints from MENTOR's mentoring standards are automated.

Challenges we ran into

Designing AI that feels approachable to people who find chat interfaces intimidating. Our answer is a narrated assistant that explains each action it takes in one plain sentence.
Keeping projects open to any field while setting clear safety rules for health topics. History, culture and nutrition science are allowed; diagnosis and treatment claims are not.
Choosing between family and institutional payers.

Accomplishments that we're proud of

A 20-screen clickable prototype covering all four personas, from voice intake to a published project page with provenance marks.
A production architecture built around verbatim storage, hash-bound approvals and multi-tenant scaling.
Health-topic safety rules worked out on a real case: a family member's interest in holistic nutrition.

What we learned

The older adults we spoke with start ambitious projects, hit an obstacle and stop, because they are working alone.
Young adults say their work matters when they feel competent, choose the work themselves and see its effect on others. These are the three needs of self-determination theory: competence, autonomy and relatedness.
Programs such as Experience Corps, SCORE, GetSetUp and Encore Fellowships show that institutions already fund structured roles for older adults.

What's next for Legado

A three-team pilot this fall, starting with our grandfather's book.
Building the first production slice: voice intake to an approved, provenance-marked chapter, on real audio.