Tagging cuisine. Drafting one-sentence venue blurbs. Writing the prose between your three plans. That's the list. Who you meet, what you do, and what's in your feed are decided by behavioral math and by you — never by a model.
not chatgpt for nightlife · not an algorithmic feed · not autoplay for your weekend
Four jobs. All of them at ingest, none of them in the request path. When you open Converge, save a spot, or vote on a plan, no model is being called — the work was done once and cached.
The names, vibes, and short descriptions you read on the vote screen. Generated in a single call after the deterministic engine has already chosen the three sets of stops — the model writes the prose, it doesn't pick the venues.
The one or two sentences under a venue name. Grounded in real review snippets, vibe tags, cuisine, and ratings — never invented. Spots that don't have enough source signal are skipped entirely rather than padded with generic praise.
For music-forward venues, the model extracts genres and energy descriptors from review text so the matching engine can tell a hip-hop lounge apart from a jazz bar. Cached on a confidence score; re-runs only when source data improves.
Restaurants without a cuisine type get one assigned from a fixed vocabulary of about 40 categories. The model can't invent new cuisines — out-of-vocab proposals are rejected and not written to the database.
That's the whole list. Each use is named and scoped, not gestured at. If we add a fifth use, it lands here in the same form.
we don't let it. a model wrote the blurb under the venue. a person — you — saved it.
9-dimensional behavioral signal: energy, social, cultural, late-night, budget, outdoor, danceability, valence, peak-time. Math, not AI.
Cosine similarity match across all users' taste vectors. Surfaces the 9 strangers with your exact taste. Math, not AI.
One-line description under each plan card. Generated per-plan from venue mix + energy arc. Claude Sonnet.
Pulling music tags from raw review text on each venue. Run once per venue at ingest. Claude Haiku.
One-to-two sentence venue summary on spot pages. Generated from review aggregation, then human-reviewed. Claude Sonnet.
Classifying restaurants into one of 40 cuisine types. Vocab-gated to prevent drift. Claude Haiku.
Ordering the venues you see in your feed. Deterministic scoring. No model in the loop. Math, not AI.
Tallying votes from each crew member to pick the winning plan. Plain count, no machine learning. Deterministic.
We think AI is useful for narrow, well-defined jobs where the cost of being wrong is low and the work is otherwise tedious. Tagging cuisine. Drafting one sentence. Pulling structure out of messy review text.
We don't think AI should replace human judgment about what's worth doing on a Friday night. We don't think it should be the front door to your social life. And we don't think shipping more AI features makes a product better.
The connection is human. The math is mathematical. The AI is administrative.
Most of our AI work runs on Claude Haiku — Anthropic's smallest, cheapest, and lowest-energy model. We reserve Claude Sonnet for the few jobs where quality matters more than volume (the plan narrative is the main one). When a smaller model does the job, we use the smaller model.
Every enrichment job caches its output. A venue gets categorized once, gets a music profile generated once, gets a description written once. The job doesn't run again unless the source data changes. We don't re-spend compute on the same answer.
We don't yet have meaningful per-active-user compute numbers to publish. Pretending otherwise would be theater. When the numbers are large enough to mean something, they'll go here.
We don't generate fake reviews. We don't ghostwrite venue claims, board descriptions, profile bios, or messages. We don't post on your behalf. We don't auto-decide where you should go, who you should follow, or what you should save.
Your taste vector is your data. It trains your recommendations — not other people's, and not anyone else's model. If you delete your account, your behavioral signal goes with it. We don't sell it, license it, or use it to fine-tune anything.
The shortest version: AI should be useful, narrow, and invisible. When it's load-bearing, it stops being a tool and starts being the thing in the way.
Spot something we got wrong, or want to push back on a claim? — email Jared