Product design · UX · Full-stack development
A closed, invitation-only web application where four founders each read a book every month, present it in a structured monthly meeting, get evaluated by an AI judge, and progress through levels, honours and a shared fund. I designed and built the product, the interface and the entire back end.
Context & problem
Four founders wanted a serious monthly reading commitment. In practice such groups collapse for a short list of reasons: choosing a book takes longer than reading one, nobody schedules the meeting, discussion stays polite and shallow, effort is invisible, and there is no consequence to skipping a month.
The group is also deliberately private and francophone, which rules out public communities and English-only tools.
The solution
01
A reader profile and a scored monthly shortlist replace the endless debate — the member still confirms the final choice.
02
Meetings are generated, scheduled across time zones, reminded without duplicates, and hosted in a video room with a calendar invite.
03
A structured question round, AI evaluation, experience points, honours, tokens and a shared fund make the effort visible and comparable.
Key features
Problem
The league is deliberately limited to four members, so open sign-up would break the concept.
What I built
An approved-email whitelist enforced at the database level, hashed single-use invitation tokens, an accept-invitation flow, admin roles, member suspension and replacement, and an audit log of admin actions.
Value
Access is closed by design and every membership change is traceable.
Problem
Members had no single place to see where they stand in the season.
What I built
Level ring with experience points, current month's book with page progress, honours grid, season leaderboard, shared adventure fund, and a live countdown to the next gathering.
Value
The full state of a member's month is readable in one screen.
Problem
Scheduling a recurring meeting across time zones by hand is fragile and easy to forget.
What I built
Server-side generation of upcoming monthly meetings, IANA time-zone handling, member slot suggestions with admin resolution, an idempotent reminder ledger so no email is ever sent twice, calendar (.ics) export, a video room per meeting, and a public scheduling endpoint that a scheduled job can call.
Value
The ritual runs itself; the admin only arbitrates.
Problem
Book discussions drift and stay passive.
What I built
A structured question/answer round per meeting: members submit questions, answers are revealed and scored, a token currency is earned and spent, special abilities are limited per meeting, and results are persisted. The rules are versioned and must be accepted.
Value
Discussion becomes a fair, repeatable, competitive format.
Problem
Peer scoring alone is inconsistent and socially loaded.
What I built
Server-side AI evaluation of submitted questions and answers through a hosted model gateway, stored per meeting, feeding a meeting report and an end-of-meeting cinematic recap.
Value
Neutral, consistent feedback that no member has to police.
Problem
Choosing the right monthly book is the hardest recurring decision.
What I built
A ten-step reader profile questionnaire, a compatibility score computed from goals, obstacles, difficulty, language and discussion value, a monthly shortlist with reasons, side-by-side comparison, save / not-interested feedback that feeds back into scoring, and an explicit final confirmation (never auto-selected).
Value
A defensible recommendation instead of a coin flip — and the member always decides.
Problem
One good book a month does not build a direction.
What I built
Thematic multi-stage paths mapped from the reader profile, with per-step state and progress, plus a dedicated path screen.
Value
Monthly reading becomes a sequence with a visible destination.
Problem
Public book data is noisy, duplicated and inconsistently described.
What I built
External search and import of book metadata and covers, deduplication by identifier, editions, difficulty and discussion-value assignment, approval/blocking, and promotion of a single Book of the Month enforced by a unique database index.
Value
Recommendations only ever draw from data a human approved.
Problem
The members are francophone; the product had to read naturally in both languages.
What I built
Full interface translation with a per-user language preference persisted to their profile.
Value
No English-only compromise for a bilingual group.
Architecture & integrations
Front end
Back end
Data
Services
Challenges & decisions
Every table carries row-level security and explicit grants, and admin checks run through security-definer functions rather than client-side flags. Execute rights on sensitive functions were revoked from anonymous and standard roles, a recursive policy on meeting participants was replaced by a dedicated helper function, and member email addresses were removed from the readable profile columns and exposed only through a self-scoped function.
The AI judge produces evaluations and signals; it never selects a member's book and never awards membership or roles. Recommendation selection always ends on an explicit human confirmation step.
Book prices could not be sourced reliably, so instead of faking them the interface states that the price is unavailable and links to a search by identifier, with the data model left ready for a real provider. Missing covers fall back to a placeholder rather than a guess.
Meeting reminders and monthly generation are keyed so a repeated or retried job cannot duplicate emails or shortlists — the correctness constraint lives in the schema, not in the scheduler.
The reading-recommendation system was layered alongside the original book tables instead of replacing them, so no historical member data was lost during the migration.
Impact & value
This is a private application for four people and no analytics were collected, so I make no claims about retention, engagement or growth. What can be shown is the work itself:
Recurring manual work removed
Meeting creation, scheduling, reminders and calendar invites no longer depend on anyone remembering.
A decision made defensible
Book choice moved from opinion to an explainable score, with the reason shown next to each recommendation.
Discussion given a format
A repeatable question round with neutral evaluation replaces unstructured conversation.
Privacy held as a hard constraint
Membership is closed at the database level and email addresses are not readable by other members.
Bilingual by default
Every screen exists in French and English with a persisted per-user preference.
Security findings closed
Permissive policies, mutable function search paths, recursive policies and over-broad execute rights were identified and fixed.
Project visuals
Member dashboard
Level and experience ring, current month's book, honours, season leaderboard, token balance and shared fund. Member names anonymised; captured in a development environment, so counters read zero.
Book discovery
Entry point to the monthly shortlist: the engine scores approved catalogue books against the reader profile — goals, obstacles, difficulty, language and discussion value.
Monthly gatherings
Meetings generated automatically per month with status (upcoming, reviews open, finalised) and admin scheduling. Each meeting carries its calendar export and video room.
Skills demonstrated
Product
Engineering
Security
UX & design
Summary
Every statement on this page is drawn from the application's own code, database schema and integrations. Where information could not be verified — usage metrics, book pricing, project timeline — it is marked as unavailable rather than estimated.