Find a cafe in Seattle area you can work and boost productivity, in under two minutes.
Try this · the buried signal
Soojin Park
Found this place during a long walk through the neighborhood. Super peaceful and the staff is so kind. I ended up sitting in the corner with my laptop for 3 hours — nobody bothered me. Outlets at every table, free fast WiFi. The latte art is also nice if that's your thing.
brew your search →
Problem: Remote workers waste ~20 minutes hunting for a cafe with reliable WiFi, outlets and a laptop-friendly policy. The answer is already in the reviews, buried under thousands of comments.
Strategy: Rejected a chatbot. Chose a search bar with filter chips that just works. Every cafe pre-tagged from real reviews and photos, every claim backed by the quote it came from.
Impact: "Laptops welcome" went from 60 cafes to 277, "ample or adequate seating" from 142 to 252. Users see 1.8–4.6× more workable options, with no transcripts to read.
Remote workers, freelancers and students lose ~20 minutes per attempt. The data exists; it is buried under thousands of irrelevant comments. The fix is to curate what is already there for the right question.
Why now? Remote work made cafe-as-office common, and AI can now read reviews the way a person does: "calm" and "peaceful" mean "quiet."
Step 1 — Survey (n=21) before writing a line of code. Tested whether the attributes I was about to build around (laptop policy, seating, vibe, hours) are what people actually weight.
| Signal | Finding | What this meant |
|---|---|---|
| Visit frequency | 67% visit cafes ≥1×/week | Realistic, not aspirational. |
| Work-from-cafe | 52% named "study or work" as a top reason | Primary use case is half the sample, not a niche. |
| Laptop-friendly weight | 86% Always/Often weight it | Laptop friendly is the primary filter chip. |
| Seating weight | 86% Always/Often weight it | Ample seating is also a key factor. |
| Hours weight | 71% Always consider hours; "late hours" surfaced unprompted twice | "Open now" option is also necessary. |
| Vibe weight | 76% Always/Often consider vibe | Vibe can vary, but for work purposes, quiet vibe comes first. |
| Pet-friendly / Parking | <15% Always for either | Confirmed dropping both from the primary chip set. |
| Helpfulness of current sources | Mean 3.57 / 5 | The laptop-friendly cafe search experience gap is real! |
"I would be interested in having a way to see what coffee shops are best to study at by answering questions like: is there Wi-Fi? How many tables are there? Can I get seating? Is there a pleasant vibe?"
— UW grad student (F, 25–29)
Step 2 — Shipped a constrained MVP.
Wall 1 · 5-review API cap → Most attributes came back blank.
Google's API legally returns 5 reviews per cafe, so volume was off the table. Meaning-extraction over fewer reviews was the only honest move.
Wall 2 · word ≠ meaning → "Peaceful" never matched "quiet."
Keyword search punishes users for not guessing the reviewer's exact word. The model had to know "calm," "peaceful" and "laid-back" answer the same question.
What I picked, and what I said no to.
Rejected · chatbot. Forces users to learn a new pattern and read transcripts just to trust a result.
✓ Selected · context-aware pre-tag. Read reviews + main photo ahead of time. Store tags with source quotes. Search is plain filtering.
Two cups, one decision: 20 minutes of manual searching → 2 minutes through pre-tagged AI.
Trust lens · invisible AI with auditable claims
Engagement lens · earning the second visit
More of each cafe written down, so more of them findable. Reading reviews for meaning, and checking the main photo as a second opinion, filled in attributes that keyword matching never reached. Measured across all 464 cafes, keyword matching versus the same reviews read for meaning:
| Attribute | Keyword matching | Reviews read for meaning |
|---|---|---|
| WiFi | 2% | 55% |
| Laptop policy | 13% | 74% |
| Outlets | 8% | 38% |
| Seating | 31% | 87% |
| Noise | 65% | 86% |
That coverage is what moves the search. "Laptops welcome" returned 60 cafes before and 277 after; "ample or adequate seating" went from 142 to 252. The dataset never grew — the same cafes simply had more of themselves recorded.
Where it stops. Outlets sit at 38%, the weakest of the five, and that is a ceiling rather than a backlog. Reviewers mention power outlets in roughly 3% of reviews; Google's own review summaries mention them in 3%; and a vision pass over 458 cafes' photos produced outlet tags for two, because Places photos are lattes and storefronts, not walls with sockets. No further scraping moves this number. It needs people who are sitting in the cafe — which makes it a product decision, not a data one.
Bars · scaled to 252 cafes
Laptop-policy coverage
+44 pts
Seating-info coverage
+57 pts
Before Needle Space
Open Google Maps → scroll 20+ results → open 5+ profiles → read ~30 reviews, photos and menus → still unsure → pick one → arrive → no outlets.
With Needle Space
Pick a location and a filter → 10+ cafes match on productivity score → take the closest → it is what was promised. 2 minutes.
The win is not the time saved. It is the certainty. Every claim on every card carries the reviewer quote behind it.
1. Check that the yardstick measures the right thing. I nearly dropped the noise tag over an agreement number. The LLM and the keyword tagger agree on noise only 30% of the time, and I read that as "noise is too contextual to tag."
The one-line version: the keyword tagger scored "cozy" and "hidden gem" as evidence of quiet, so it called 301 of 464 cafes quiet and never once said loud — it was measuring charm, not sound.
Its keyword list put "cozy", "small cafe", "intimate", "tucked away" and "hidden gem" in the quiet bucket, each worth +2. Those words describe ambiance and appear in most cafe reviews, so almost every cafe tipped into quiet before a real noise word was weighed. The LLM reads the sentence instead of matching the word: of its 88 quiet tags carrying a quote, 87 cite an actual acoustic statement — "quiet enough to actually hear the person sitting across from you." Across the 197 cafes where the two disagree, the LLM cites supporting evidence 89% of the time. Low agreement meant my baseline was wrong, not that the feature was unbuildable.
What I did about it: 28 cafes were still displaying "Quiet" on keyword evidence alone, because the app fell back to the keyword tag wherever the LLM had none. Those now show nothing. An empty space is honest; a confident wrong label is not.
Agreement with a baseline is not quality. I replaced it with two measures that need no ground truth — how often the tagger has to answer unknown, and how often it can cite a quote for the answer it gave — and gated the monthly pipeline on those instead. A metric that would have talked me out of a working feature is worse than no metric.
2. Check with the possible options in the beginning phase. I spent hours building keyword-based filters before realising they miss the context of a review. Exploring the options first would have saved all of it.
The constraint forced a better product. Five reviews per cafe meant I couldn't lean on volume or even web scraping due to platform violation. I had to make the system logic that actually can increase the accuracy. The user never sees that work. The user wants to go to the best cafe to work, and searching for the cafe should be their minimum efforts.
⠀
Live prototype
Live at needle-space.netlify.app ↗ — embedded below.