Dine Guide

Answer the timeless question: “Where should we go for dinner?”

This app aims to replace Yelp by offering a faster, more social, and more personalized way to choose restaurants and dining experiences with friends.

Rough draft of dream UX: https://chatgpt.com/canvas/shared/67f98d74d7d48191a517abc3dd923de3

Funny note: This is pretty close to one of the canonical examples of a tarpit idea, as described by Dalton & Michael at YC. The original video seems to have been removed from YT, but here’s V2: https://www.youtube.com/watch?v=GU9iT7MW0rs

Why I care

See Local connection. Dining in groups at restaurants is a more narrowly scoped activity that checks a lot of those boxes.

Status: Prototyped and Parked

I worked on this from April through July from a few angles. I focused primarily on the discovery part, since loading up a database with AI analyzed restaurants seemed like a necessary precursor to the other functionality. This was also a good opportunity to play more with agentic development. A few of the technical outputs of the process:

• Zero hand written software. This isn’t appropriate elsewhere, but a fun experiment for a POC here.
• Stack
  • Rails back end
  • React front end
  • Custom CSS design (wouldn’t do that again with LLMs)
• Functionality
  • Google oauth
  • Google maps places database seeding via orchestrated LLM prompt frameworks + playwright MCP
  • Multi pass AI analysis and caching of review text and images with OpenAI API
  • AI summarization of independent AI analyses for performance
  • Dynamic data collection and seeding in new venues
  • Natural search intent parsing with Open AI, with location hybrid inference with browser location.

Technically, I learned

• About the costs / speed of different OpenAI models
• The scaling / cost issues with AI analysis of this quantity of content
• How much more effective it was to analyze images than simplistic text reviews - so much more information can be inferred

Playing with the POC, I learned / concluded a number of things

Data

• Crucial details like restricted diet menus are generally not present. Things like open hours are not always reliable either
  • These make the hopes of decreasing bad recommendations impossible on current data
  • A deeper AI analysis did not deliver a 10x experience for me
• From playing with POC as well as follow up conversations, it seems the best restaurants are extremely obvious through the volume of reviews and other signals. Getting to a confident recommendation for something with few reviews is far harder.

Key insight in here, validated with other people:

• Most people have an internal ~4.5 star / 500 review heuristic filter with restaurants
• However, it only takes one recommendation from a friend with roughly similar taste to justify a visit.

This dynamic is possibly also why marketing on these platforms is challenging. Newer restaurants are outside that low risk solution set.

Another insight in that 4-5 star range: There are a ton of restaurants that are great for specific things. They earn some bad reviews from people going in with incorrect expectations. The new Google AI features start to address this, since (at large review volume) it’s pretty easy to infer recommended occasions to visit, things to order, etc.

User experience

Finding the best restaurant seems like a simple goal, but it’s poorly framed. There’s quite a bit of context required about both the customer and venue to make a meaningful improvement over the current platforms

A possible reframing of the goal is finding a reason / activation energy toward an option. Move beyond decision fatigue. That’s a place to iterate and understand user motivations. What makes people take a risk on a place and be pretty happy regardless of outcome?

Business opportunity analysis

Building on top of Google’s data for prototyping purposes was illuminating. Even scraping everything users can see (which is more than their API exposes), I wasn’t able to transform that data (sparing no AI API expense) into something that delivered a 10x experience in matching me to something new with a compelling premise.

Obviously, building on Google’s data isn’t a viable approach for a production app either.

Likely requirements for success in this problem space

• Need to own your own review / evaluation data for a restaurant
  • Which means solving the chicken / egg problem to get to sufficient data in at least one region, then repeat
• Create a new user mental modal around recommendations that’s not based on review count & stars
  • To help the launch velocity and have a chance to build a 10x experience overall
• Collecting more restaurant data - accurate hours, menus - including dietary restrictions, specials, events.
  • Enable very tailored recommendations to the customers who care the most

https://beliapp.com/ is the closest thing I’ve found. Personally, I don’t think the ranking of dissimilar restaurants makes sense as the core rating/review mechanism. Friends have said the same. Its slow roll out after years of development hints at not figuring out a compelling formula here.