I built an AI Wine Sommelier - and started a tiny rebellion...

As part of my ongoing series of seeing if I can turn my hand to literally anything using the awesome power of AI, here I document my latest travails on chardonnAI - my AI-powered wine sommelier service.

Introducing ChardonnAI - your own personal wine sommelier... right in your pocket.

ChardonnAI

For those who missed my first blog post on this, I've been having some fun designing, building & now launching an AI-powered wine sommelier. Naturally, AI has been my invaluable companion here - chatGPT Work to be precise, and largely done using Luna 6, with the occasional stretch into Sol & Astra.

I won't proclaim this is revolutionary - indeed a brief search on the app store yields at least 50 AI sommelier apps - but it has been great fun building it. Myself and GPT Luna have effectively been this sort of duo over the last few days, bouncing ideas back & forth, tweaks, opinions, perspectives - and hey it's been fun learning more about wine at any rate.

Given so many wine sommelier apps exist, I've decided to market it as the 'anti-wine-app' wine app (hehe..) - it deliberately plays into the fact that all these other wine apps have wine catalogues of 250k+ wines... let's cut the nonsense and get to the short-list of the most popular & most reasonably priced wines that someone will actually be able to buy and find on a shelf somewhere. It also just happens to make my life easier in terms of curating said wine catalogue... ;)

The premise is pretty simple - you land immediately on a chat interface with "ChardonnAI" and spend a little bit of time conversing with her in terms of what wine you may be after, a particular meal to cater for, a particular taste - and ChardonnAI will go fetch the perfect match from its curated catalogue of 250 of the most popular & reasonably priced wines. It gives a final justified recommendation replete with Body, Acidity, Tannin & Sweetness measures, a price range, a 'did you know' section, auto-links to take you to the most popular UK retailers to buy the wine, and gives you the ability to vote whether you agree or not.

It also learns - so the more you converse, the smarter ChardonnAI will get at matching you with the wine you'll love. Nifty feature though that memory is only local to the specific device at the moment. I'd have to figure out user logins to improve that...

The full 'Wine Catalogue' is detailed on a separate page and fully explorable with click-through detail per wine, there is a 'Palate IQ' page that gives ChardonnAI a helping hand in understanding what sort of wine you like, and also a Test your Wine Knowledge quiz with ChardonnAI herself.

So what did I learn and what were some of the roadblocks:

• Vibe-coding this up with GPT was actually fairly painless. It did take quite a bit of re-prompting though - indeed, I sort of got into the rhythm of just having a very protracted conversation with chatGPT through the whole day while I was going about other stuff. She'd prototype the html - I'd ask for fine-tunes. It took quite a while as I was probably using a slightly underpowered model for this - but I'm increasingly finding that going up to Sol high, or (god-forbid) Astra tends to murder my daily allowance within about 15 minutes. So yes, it's faster to create, but then you need to wait until the next day for allowance reset... so it's longer in the long-run.
• I had to do a fair amount of testing & tweaking of the persona and personality of ChardonnAI. In my last blog I explained how she became depressive because she couldn't match the right wine to my chicken madras - after that she became too rushed - demanding to know my budget before even really asking what wine I was after. I then tried dialling up the sommelier snobbery but went too far. Finally I think I've reached an informed, lively & brief sort of tone & personality - but you be the judge.
• I had to set up a proper API through the OpenAI Developer platform for model calls to support the conversation - pretty painless. It's great that you can see exact token usage and spend. I loaded $30 into my API account hoping that would be enough - turns out to get to this point and all of the prompt testing I've done, I've spent just $0.05. Crazy cheap. The risk I run is if this app does actually blow up and get popular, I'll be funding the API calls unless I decide to monetise this in any way. But it feels too early to monetise - I'd rather just keep it free and see if people actually like it. I dunno, isn't monetising an anti-establishment wine app a bit... 'establishment'??
• Wine images are still a bit of an issue. Image rights was a LONG conversation between GPT and me - it felt at times like I was conversing with a lawyer. Basically (unsurprisingly?) you can't just scrape any old images from retailers or risk the wrath of litigation. There do exist some free-to-use libraries but it's a lot of searching and hassle. In the end we managed to (legally) get good-quality full wine profile shots for maybe 20% of the wines, for the rest we decided to use AI-generated shots of the wine set against a representative background of the country of origin of the wine. Neat.
• Domain was cheap - £7.99. I wanted to get chardonn.ai (everyone knows all the cool websites are .ai now) alas it was taken. I'd need to shell out £80 just to hire a 'domain broker' whatever that is...

Overall - honestly, it's been great fun. I have absolutely no right whatsoever to be able to do this and take this concept E2E into a launched product - if it wasn't for AI helping me. But I guess that's the awesome world we live in nowadays...

So yes - please do give ChardonnAI a go. Tell me what works / what doesn't. If you like it / love it, tell your friends! It's all about the anti-wine-app wine app! Viva la revolution!

Next up I'm going to see if I can become an (AI-powered) iOS developer and make the ChardonnAI iPhone app. In fact I'm already knee-deep in it...

Ciao