Hi all! AskMaps is a side project of mine. You chat with an LLM (Claude, GPT, or DeepSeek, your pick) and the places it mentions show up on an interactive map as you talk.The backstory: about a year ago I took a few spontaneous trips through Europe and didn t want to spend ages researching beforehand. I asked Claude instead, and it turned out to be a really good travel advisor. For most destinations it has a ton of good tips. The only thing that sucked was the UX. Copy-pasting place names into Google Maps one by one is tedious, and it s even worse on your phone once you re already there and just want to see what s nearby. So I built myself something that works the way I wanted, and figured it might be useful to you too.I know Claude and ChatGPT can sort of show things on a map themselves, but it s not really usable. It s buried in the chat history, it isn t interactive, and you can t see where you actually are right now.I started building in October, had a first version in November, and have been dogfooding it hard on my own trips since. It s at the point now where I rely on it. It factors in the weather and your current location, and it ll warn you before you book a flight somewhere you shouldn t (thanks to their training cutoff LLMs will happily recommend a sightseeing trip through the Persian golf region). You can also group several chats into a project so they share context across one trip.There s a demo with some example conversations you can click through, or you can sign up and try it for real, you get some free messages. I had to put it behind a signup because leaving a chat window open to the whole internet would kill my wallet.The two biggest challenges I faced were geocoding reliability and LLM costs.For geocoding, I didn t just want pins, I wanted real geometry wherever possible. A pin labeled Le Marais is fine, but if I m standing nearby I d rather see the actual area than eyeball it. OpenStreetMap is the obvious source, but you have to fight it. Ask for a neighborhood and it ll happily hand you a random bus stop, or a POI with the same name on the other side of the city. So the geocoder grew into a whole system: multiple sources of truth, custom ranking and automated conflict resolution via LLMs.With the LLM, I started on Sonnet 4.5 because it s great, and it (then 4.6) stayed the gold standard for a long while. A lot of the work went into keeping it cheap: maximizing cache hits and cutting tool-loop iterations, since with Claude every extra round trip means a fresh cache point and more money. Then DeepSeek cooked with V4 Pro. After a lot of testing and tuning it came out roughly on par with Sonnet at about 20% of the cost, so it s the default now. Works great, and keeps credit usage down for users too.Happy to answer anything!