Interest comes first
Events are ordered by stated interests, not by volume or recency alone. People see what fits before what is merely nearby.
← Selected Work
Working build · Kauaʻi community data
Not-for-profit · Community platform for Kauaʻi
Find something worth showing up for.
MatchBook is a not-for-profit community platform for the island of Kauaʻi. It gathers the island’s events and groups, matches them to each person’s interests, and turns the result into a plan. It is free for residents and visitors.
Not-for-profit
A community project, built to serve the island rather than a market.
Free of charge
No fee for residents or visitors — at launch and after.
Launching soon
Public release is weeks away.
Working build · Kauaʻi community data
The front door: what is happening nearby, this week, and what to join next.
283island events
24community groups
9interest topics
7core surfaces
1island
The current working dataset: island events, groups, and venues used to validate discovery, matching, and planning ahead of launch.
01 · The product
One place for the island’s public events.
Kauaʻi’s community life is announced across county calendars, library pages, radio calendars, flyers, and social feeds. No single place holds all of it, so events reach the people already watching the right channel and miss everyone else.
MatchBook consolidates that information and matches it to each person’s stated interests. The result is a short list of events that fit — with the schedule, the location, and the group behind each one.
The questionHow short can the path be between hearing about something and attending it?
Planning is part of discovery
Saves and RSVPs land on one calendar, so deciding and scheduling happen in one step, not across two apps.
Organizers use the same tools
Groups create and maintain their own events inside the product, through the same interface residents use.
02 · How it’s used
From interests to attendance.
Interests are set once. Everything after runs in one loop.
- 01Set interestsChoose the topics that fit how you spend time
- 02BrowseSee events and groups, ordered by fit
- 03DecidePass, save, or join — one action per card
- 04ShortlistSaved events collect in a single list to choose from
- 05ScheduleConfirmed plans land on the calendar and agenda
- 06AttendThe week is planned; the rest happens in person
03 · The screens
The product, screen by screen.
Each surface handles one job: discover, decide, or plan.
04 · The brand
One mark, applied everywhere.
A drawn wordmark, an illustration set, and motifs for the core interest topics.
The matchbook is the organizing object: small, familiar, made for starting something. It appears in the wordmark, in the match illustrations, and in the interface’s empty states.
Each topic carries its own motif — music, food, outdoors, community, books, making — used consistently across cards, calendar entries, and filters, so categories read at a glance.
05 · The platform
A deliberately small platform.
Interfaces read live data. Every write passes through validated, permissioned server methods.
Experience
React 18 SPASeven surfacesCalendar viewsSwipe matching
Server boundary
Meteor methodsSchema-validated writesScoped publicationsRole-based access
Data model
GroupsEventsProfilesMembershipsEvent links
Island dataset
283 events24 groupsReal venuesWeekly rhythms
Live interfaces
Calendars, feeds, and group pages update as data changes — no refresh, no polling logic in the interface.
Server-validated writes
Every change passes through schema-checked methods. The client cannot write to the database directly.
Scoped permissions
Publications and role checks limit each account to the data it is allowed to see.
Organizer tooling
Group management, event creation, and administration run in the same design system as the public product.
06 · The road to launch
Two systems are in development for release.
An automated register of the island’s public events, and a tiered recommendation system. Both follow the same rule: automation gathers, people decide, and nothing ships until it works.
In development · Automated ingestion
The community event register
A single TypeScript service will collect public events from the island’s institutional calendars — County of Kauaʻi services, the state library system, community radio, festivals, and the community college — on a schedule, through per-source adapters, without paid APIs.
Collection and publication are strictly separated. Automated collection produces evidence; a human review decision is the only thing that changes what the public sees. Every published event carries provenance back to its source, and anything unknown — a price, a venue detail, an accessibility note — is stored as unknown rather than guessed.
The central ruleCollected data creates immutable source observations. Only an approved review decision can modify a canonical public record.
First-release sources
County of Kauaʻi
Hawaiʻi State Public Libraries
KKCR Community Radio
Kauaʻi Festivals
Kauaʻi Community College
Kauaʻi ADRC
Coconut Wireless
Kanu Hawaiʻi
- Sources
Public county, library, radio, festival, and college calendars — each with its own adapter and each site’s crawl rules respected.
- Snapshots
Responses are hashed and stored once; unchanged material is never reprocessed.
- Observations
Each record becomes an immutable observation with evidence pointing into the snapshot.
- Candidates
Normalization, validation, recurrence handling, and duplicate detection under fixed policies.
- Review
People approve, merge, or reject. Similarity can suggest a duplicate; it can never authorize a merge.
- Public record
Approved changes publish atomically, with audit history and provenance intact.
In development · Tiered rollout
The recommendation system
Recommendations will power the swipe deck and the main feed from one server-side service. The system is tiered: the launch tier is designed to work well with sparse data, and every learned model above it stays behind a feature flag until it beats the tier below on chronological evaluation.
The foundation is data that tells the truth. Interested, saved, going, and attended will be separate actions in an append-only history. Impressions will be recorded only when a card is actually seen. Distance will never be estimated without real coordinates, and unknown will stay unknown.
Every recommendation explains itself
“Similar to events you saved”
“From an organizer you’ve enjoyed”
“New in your area”
“A little outside your usual”
- T0Truthful foundationIn development
Append-only interactions, honest impression logging, normalized event data.
- T1Graph-aware hybridLaunch model
Content similarity, time-decayed graph affinity, schedule and distance fit, diversity re-ranking, controlled exploration.
- T2LightGCNGated
Collaborative graph learning over users, groups, organizers, venues, and topics.
- T3Heterogeneous graphGated
Inductive models that can rank events with no interaction history.
- T4Temporal graphGated
Weekday, seasonal, and momentum patterns learned over time.
Promotion requires beating the previous tier on a chronological holdout, then in shadow mode, then in a limited live test. Otherwise it stays off.
07 · The standard
Community information, treated as infrastructure.
MatchBook treats what happens on Kauaʻi as shared infrastructure: free to use, collected from public sources, and governed by human review.
The same standard is being built into the engineering — server-side authority, append-only history, and systems promoted only when the evidence says they should be.