Technology
The intent graph.
Swink turns what people say they want, and who they actually choose, into a graph of real-world relationships that gets more accurate with every event.
A graph built from intent, not claims.
Professional networks record who people say they know. Swink records something different: who wanted to meet whom, why, and whether the conversation continued.
Every event circle produces three kinds of signal: the purpose statement a person writes, the mutual opt-in decisions they make, and the conversation behavior that follows.
Graph definition
The signals that compound.
Pipeline
From purpose to possibility.
Purpose statements, opt-ins and chat signals.
Convert text into vectors.
Surface candidates inside the event circle.
Score intent fit and mutual-interest likelihood.
Update edge weights from outcomes.
NVIDIA accelerated computing
A conservative architecture, clearly stated.
| Component | What it does | NVIDIA technology | Status |
|---|---|---|---|
| Text embeddings | Turns purpose statements and profile summaries into vectors for semantic matching. | NVIDIA NeMo Retriever text-embedding models, served as NVIDIA NIM microservices. | In development |
| Vector search | Finds nearest-neighbour candidates within an event circle in milliseconds, including at peak join times. | NVIDIA cuVS (RAPIDS) GPU-accelerated vector search. | Planned |
| Purpose-statement assist | Helps users turn a rough goal into a clear one-line purpose statement. | LLM inference served with NVIDIA NIM / TensorRT-LLM. | Live in beta, using current provider |
| Graph analytics | Computes edge weights, communities and re-engagement patterns across events. | NVIDIA cuGraph (RAPIDS) on GPU. | Planned |
| Ranking model | Scores candidate matches for intent fit and mutual-interest likelihood. | GPU training with PyTorch on NVIDIA GPUs; graph neural network option via cuGraph-PyG. | Planned |
| Model serving | Serves embedding and ranking models with low latency and batching. | NVIDIA Triton Inference Server with TensorRT optimization. | Planned |
Why GPUs matter
Events are bursts of demand.
A networking circle is busiest in the minutes after doors open, when everyone joins, writes a purpose statement and expects relevant people immediately. Embedding many statements at once and searching across a whole room needs parallel compute.
Across events, the intent graph grows into millions of relationships. Computing edge weights, communities and ranking features over that graph is a natural fit for GPU-accelerated analytics.
Swink is applying to NVIDIA Inception to access accelerated computing, technical guidance and training as we scale from beta events to dense, repeat-event communities.
Quality & responsibility
Intent data is personal. We treat it that way.
We evaluate matching on outcomes people actually choose: the rate at which suggested people are opted into, the share of matches that become conversations, how many conversations continue after 24 hours and seven days, and how often people reconnect at later events.
Purpose statements are visible only within the relevant event circle and after confirmation. People can edit or delete them.
Next step