SOCIETY · UWBHACKS · 24-HOUR BUILD
EVENT NETWORKING GOES ON-DEVICE
You walk into a room full of strangers — Affiniti figures out who you should meet, without a single cloud call
BOTHELL — Events and hackathons put the right people in the same room, then leave them to find each other through an alphabetical attendee list and awkward small talk. The valuable connection — someone who shares your niche interest and is free to talk right now — stays invisible. Affiniti, a roughly 24-hour MVP built at UWBHacks on two iPhones and one local server, makes that connection computable.
When an attendee RSVPs, a local language model produces a semantic summary of their profile, which is embedded into a 768-dimensional vector; their stated intent for the event gets the same treatment. At match time the system blends profile and intent similarity 60/40 through pgvector cosine search over an HNSW index, with a skill-diversity constraint so a builder does not draw four ML engineers in a row.
HOW IT WORKS
For every match, Gemma 4 — running on Apple Silicon via mlx-lm — writes a personalized icebreaker grounded in what the two people actually share. No generic prompts, and no data leaving the device.
Inside each event, the SwiftUI app offers four tabs: a ranked discovery feed, the top four matches with their icebreakers, bookmarks, and a live status broadcaster — "Open to Chat," "Deep in Work" — showing who is actually available at that moment. The async FastAPI backend, with SQLAlchemy and Auth0 JWTs, served both phones over ngrok.
LESSONS FROM THE FLOOR
Ranking on profile similarity alone produced echo-chamber matches; blending intent plus the diversity constraint made recommendations feel useful rather than redundant. And wiring real devices to one local server surfaced JWT and threading issues the simulator had hidden — where, the builder notes, hackathon demos are won or lost.
Source available → github.com/SNBest1/Affiniti ↗
CONTINUED: Caregiver alerts ride event-driven AWS, Page B2 →