From the event: Mindstone Miami June AI MeetupWhat I Learned Building With LLMs
View event

What I Learned Building With LLMs

Introduction

Hello, everyone. How are you today? Tonight? I have a question for you.

So we talked a lot about what we do, but you may really like this presentation. How many of us have actually created videos and photos using AI? How many of us?

How many of you guys have done that for your business with social content? How many? How many of you guys have done that with your own avatar?

you have yeah how many of you have heard of Higgs field Higgs field what a Higgs field yeah yeah yeah yes Higgs field all right so we're gonna talk a little bit more about that and if you guys have any questions please let me know I consider

myself like somewhat technical I did do this for a non -technical audience but but sometimes I'm told that I'm a little bit more technical than I think, so if I say anything or have anything and you have questions about it, please just stop me, okay?

Who I Am and What We’re Building

So my name is Alana and I am going to talk to you about what I learned while building with LLMs.

A little background about me, I just graduated from the University of Miami with my master's in data science and concentrating on AI.

I absolutely adore fashion I've been in almost every job of fashion design styling visual merchandising and everything so I like to stay up to date with like the business side of fashion the marketing side of fashion my undergrad is in business so I've always been interested in not just the

creative side but the business side when AI came out it brought me down the rabbit hole of how can I figure out why products become popular like what makes the people adopt different products is it because they love a founder is it because they are just loving the content that people are putting out is it

because the actual product is good quality like what is the formula and how How can I find that formula fast for any brand that I want? So I got into data, AI, I love fashion.

From Fashion to Data Science to a Marketing OS

I started my master's program at U .M. And I started making this tool and I was submitting it for projects and talking to people about it

and people were like, this is amazing. We want this, so I built a full intelligent marketing operating system.

Today’s Focus: Creative Workflow + Community Dashboards

And today, we're not gonna look at the full system, But we're going to look at the creative workflow of the system and some dashboards of how to look at your community organic content.

Because as we always hear, the creator community is booming. We want to spend the least amount of money as possible. So how can your organic content and founder -led influence really get some better information on your marketing strategy and on your sales goals?

Any questions? All right.

What I Learned Building With LLMs

so first just a little bit what I learned every LLM has a strength we talked a little bit about this before so you need to know what you need and you to know each LLM strength unfortunately and fortunately that changes all the time so you want to keep abreast I watch different YouTube channels and I play

Picking the Right Model for the Job (A Practical “Big Three”)

around with all the tools all the time so if you can't afford it know what's good my this is my Avengers this is my big three anthropic for research and and strategy.

Google Gemini is great for UI design. I do a lot of UI and UX design for a company in Coral Gables.

And OpenAI is great with natural language.

So for me, if you have these three, you can do anything and build anything. And this is how I work with them.

A Human-in-the-Loop Workflow: Direct, Verify, Decide

I love automation. However, I am a creative at heart, even though I want something that does take less time.

so I like to be the director all right so I direct it I don't ask it once you start working with these LLMs you start putting in prompts it can take you wherever it wants to take you so make sure that you're always the director

make sure that you make it site information it likes to hallucinate especially with updates and models make sure that it stays on track and you want to make sure that you give it context with every input every time you input give it all the context pour out everything that you need ask for an

understanding of information right so say you want to generate an image you want a camera angles facing forward if you're not familiar with it you're gonna need to ask like what are my options if I want to face for it can't camera angle so you actually have an informed decision when you come up with that

final prompt and then you want to read the full output so that you have an understanding of where it went this is where a lot of people go wrong they're not reading the full output they're using too much compute and then the lm just goes sideways so again this is

my favorite stack cloud chat gpt i use aura a lot which is has anybody heard of aura does anybody use aura it's my favorite i know people like figma you can import things from figma but you can use aura i love to use gemini and aura i use claude when i need something that's a little bit more code heavy.

I use cursor for the back end after I do my UI. And I love Hicksville because it has all of the models there. And my favorite, of course, is the big nano banana pro.

Nobody better. And C dance. Okay.

So just to reiterate, you direct, you route to the right LLM, you verify and decide where you want to go. That is my workflow. That is the best way I've

learned how to work with LLMs. I know people love automation. They'd rather sort through the slop, but this is the best way to get some good results,

actually learn along the way, and be cost efficient. Okay, any questions? All right,

Demo: Prompt Builder and Creative Production Flow

so I'm gonna move to a short demo. I recorded a video because, honestly, mainly Mainly because I can't spell and nobody wanted to see me do that up here.

So this is what, actually let's start here.

Reducing Decision Fatigue With a Prompt Builder

I do a lot of UI UX, I do a lot of websites, I do a lot of photos with real footage and then insert avatars, styling and videos with real footage and insert avatars. so I had to build myself to prompt builder this just takes away a lot of decision fatigue a lot of the things that I normally have to decide anyway

it's just super quick here a couple of buttons so I can have a concept I insert the concept here like view of a girl right and it starts the prompts over here or I can generate it from an audience I can enter a Gen Z World Cup promo for energy drink and then it'll insert the information here you can

enter their brand aesthetic I can choose the colors and I can enter the exact color of the brand the highlight color different content type UGC all the important things that you want to build the creative direction the camera angle

the lights the mood text or no text all of the information I can copy it and clear it and then I will show a demo video of me actually doing it so it's a little faster for yeah okay so I copied in an idea already chose the brand brand aesthetic.

My idea was so heavy and busy, honestly, that I didn't need to add a lot of information here. A lot of this is for if I have a company that doesn't really have that much of an idea and I can build something out for them, or if I just want

to demo what they do. I missed a little part here. I typed in random words, and then just

just to, if I want to add something that the prompt didn't have, and then I can clear that only what I typed and it can take me back to the assembled version that I did on the left side that the prompt created and clear only what I typed without clearing the entire prompt. All right.

Claude + MCP + Higgs Field: Generating Images and Video

So I got my prompt. I took it into Claude.

How many of us have heard of an MCP? Did I already ask that? Okay.

So if you haven't heard of an MCP, it's basically just how to AI models can talk to each other Higgs field has MCP with Claude and this project is connected to Higgs field so that it can run and create videos and images as much as I want to so I'm gonna show you a process that is heavily human

input but you can say I'm running an agency and I have five products that I need to deliver on I only have like my packages eight photos right only need eight photos i can open this project open five separate chats at the same time input the product say i need a hundred images and i need them to have these type of categories this type of background and outlook walk away come back in a few hours i'll have a hundred images for each product and then i can sort through and decide the best ones and then deliver it to my client

so how can we not do that under like cloud design because i found that doing it in a regular cloud on any kind of design work it just it just sounds good yeah yeah i well how come i'm not generating the images under cloud design yeah yeah because the model nano banana pro is just so much better

so hicksville gives you access to all of the models and you can play around with the different ones but nano banana pearl is the best for images and for me and c dance is the best for videos and And then in Hicksfield, if that changes, which, like we said, it always changes, you can just tell Claude in the chat, actually use something else. Use Gemini or VL3.

What does MCP stand for? I don't know what it stands for, actually. That's a great one. Model Context Protocol. Okay. Model Context Protocol. Thank you.

I'm wondering, so do you have, like, a plan that allows you so many tokens with Hicksfield? Yes. In Higgs Field, there are several plans. There are so many different plans.

It depends on if you're more image heavy or video heavy. You can go in there and choose your plan and

then when you connect them, it doesn't, your CLAWD credits are one separate thing than your Higgs Field credits. So you can plan as much as you want in your project, in your chat in CLAWD and then once you're ready to go, you will call Higgs Field and it'll only

do like nano banana pro is four credits per image I don't only use that four credits while you're planning things out. Do you find Higgs field to be like expensive for token burn?

I don't it is expensive for video imaging but it's everywhere it's expensive yeah but if you really plan out the prompting then then you have less iterations and it's really affordable.

And do you need to like trim your context window as you like refine or does it do it for you? Yes, so what I created with my prompt builder, I did research with Claude on how nano banana

likes to be prompted and CDance likes to be prompted and I created a framework so that the prompt outputs puts based on how Nano Banana Pro has put the information they put out on how it likes to be prompted.

So the way that it talks is the way that Nano Banana Pro likes to be talked to. And then once the prompt gets pushed out you put it back in Clawed and this is what it's going to show now and it cleans it up even more with the reference of the actual product and then it'll give you like a really good prompt.

so do you do you have like a method for storing context for memory no and that is what I'm building building that's what my operating system is one of the major things is building is just my own knowledge of how I like to create in my studio and it's a creative content studio within my marketing operating

system yes but it's a lot of compute so it's more of just I have the bones of how to collect it but I haven't tested if it's actually understanding anything outside of the project outside of my cloud project understanding what I like does that make sense yeah no problem no Higgs field is a separate it's a

separate subscription a separate company but the MCP allows them to talk to each other and I actually built this product from scratch in Higgs field this isn't a a real product no that's just another chat no what it takes feels the creator

of nano banana what is Google yeah yeah sorry Higgs field is just at home where all of the models live and then it gets more specific so like if you want UGC content user generated content they've created their own logic where when you you put in a prompt it knows to only give it give you UGC content and things

like that. So what would you tell them between Higgs field and HuggingFest? Is there a difference between these two? HuggingFest?

I'm only familiar with HuggingFest as far as databases. Higgs field is an image generating, a video and image model generator and then it has they have their own model generator

that's not that great but they house all the other models yeah so yeah I built this with nano banana Pro using Hicks field you getting a bottle off of Alibaba and getting some white label information and built the real product

and then now I'm going to build an actual like social content post so I pasted the prompt from my prompt build on here I'm telling Claude to enhance the prompt understanding that this is the product that needs to be in here surprisingly I didn't misspell anything

this is really raw I didn't like look for good things so there's a few hiccups that we have in this that's actually gonna make it kind of fun so just a pause here you can have this like I said completely automated but I don't allow it to do that some people have separate laptops where they run things like this

and they let Claude and other AI models just run free but I don't do that so I'm allowing it at every step so every time hey Claude wants to talk to Higgs field and click basically I have to allow it and it's asking me to upload the image again this is because now it's in Higgs field so Higgs field wants to actually

see the image and then it's gonna spit out one of the ugliest product photos you've ever seen and this is because the data of what the prompt it was so messy I did not read the output I was like I just want to see how it goes for the video so I

said okay let's get serious and then I wrote a real prompt and then it gave me this cool little photo which looks like all of the photos for energy drinks that that are on these websites right now.

And so I did one more, if this will play. And this was really easy. It's so much easier to do it in the chat and in the project. I was like, ew, this is ugly.

Just put it in the desert with ginger next to it and a lightning bolt. And it redid the original prompt, and it did that within seconds instead of going back and forth

between models. OK. So we have. Excuse me.

So I see you're using Opus 4 -8 High, that you were, just to do that? No, I probably just had the tab up. No, I'm only using Claude in Higgs field. Sorry to be confusing.

I'm just using Claude in Higgs field, and within Higgs field, I'm using Nano Banana Pro to create these images. Are you doing anything with the Anthropic Project model? yeah this is its own project so I can show you later but it is best to create

a project that's just Higgs filled MCP and then when you want to generate images you open a new chat in the project and so that is how my what I was saying earlier is how I'm gathering like a report on how I like to design images and videos because all of the information is in one project okay so

Yes.

Real-World Iteration: Quality, Safety Filters, and Fixing Prompts

this is another prompt another idea I had for another product I made this one's a little bit more fun because I did was fun for me because I did two images more spelling my kryptonite and let's see so I'm asking to enhance this prompt and I

have two photos one of them is an avatar technically this is a person I created in Higgs field and did a character sheet and the second one is a product that I created in Higgs field and designed it and I got a little bit more inspired added some more information misspelled some things again had a prompt approved

the prompts and then it gets interesting because in the prompt I wanted the serum the hair product the skin product to be in a outside beach area where the girl had on a bikini and her hair was wet and it said please ma 'am let's be respectful this is not safe for work and I was like what is happening I said try again

because sometimes that just happens and it said no ma 'am so in the chat again this just makes it so much easier I said this keeps saying not safe for work what's the problem it immediately fixed it and it redid the prompt and then it it pushed out a product photo with a girl with her hair wet. Maybe you should try it.

Grok would have let it happen. Grok would have definitely let it happen. And so then you have a product photo. And again, I want to reiterate, that is not a real product.

I got some random bottle off of Alibaba, created the product, the logo, everything. That is not a real person. So you can get really fun and creative with this.

Beyond Product Shots: Avatars, Styling, and Mixed Reality Content

now this is just kind of wonky but you know why not I wanted to show you guys some other things that you can do you can also style like I said I love fashion so this is from my favorite clothing store forward made an avatar put the outfit on the avatar which is a rendition of me something sometimes I look like this got a vintage chair desk that I really thought was cool put a macbook on it and then had an intro video to my website none of this is real then i made myself

a pencil sketch and then i said how crazy can i get can i make a fake world and make me walk into here the audio is it working let's see in a fake road and then i put myself in the real world that is real footage from flow miami and put my pencil sketch in there so you can get really fun fun with these things, put out some cool content, and then collect the data.

Measuring What Works: Organic Content and Community Analytics

The main point of me really getting into content creation is I wanted the data. I want to know what makes people dry, what makes people want to buy.

Like I said, is it the founder, creative director? Is it the content? Is it the product?

Dashboards for Reach, Sentiment, and Revenue Impact

so this is dummy data but all the brands are here and it can just tell you how much are you actually getting from each of your social posts and this is organic which is the game changer because there are a lot of companies out there who do this first paid ads but people don't want to pay for ads the creator

community is really big and everybody is focusing on a loyal community and a lot of founders are having a personal brand and so I was interested in how does this affect the actual sales of these companies so it has the revenue per 1k

reach did it clear the line after being organic for the first three days was it going so viral that you should actually put money behind it how many people are saving it just the sentiment how is the sentiment going are they liking you are Are they loving you and how is that affecting the sales?

Dark Social: Detecting Shares You Can’t Directly Track

And then the dark social, which is really interesting because of course we can't scrape a lot of things like Instagram and stuff, but you can see when someone pops into your website directly to a product, you know somebody sent them that link. Or they go straight to a promo page that's buried inside, but they didn't click it from

your website or social media or something, you know that somebody sent it to them. so how shareable outside of the internet is your actual content and it's just broken down by platform what's performing well and how much revenue is there estimating to come from it and then you can click on the actual content

page and then content that you posted it and see a breakdown of that individual content so you actually the prompt that you build you can give it to the customer right or you have to use it i can give it to the customer but everyone has just wanted

the actual final product basically what i'm saying here you can quickly or put someone in the house let's say someone you have a house for sale so you can create like a virtual look wheel how you're gonna be in the house this is the good thing about the house yes make that person you know live in in the house why not living in the house yet exactly and then present that in the social

media and make that that sentiment that emotion exactly basically that's what you would do yes you can there's so many creative ideas but yes you can definitely do that and then once it's out there see how that's actually how can that help you with your next marketing campaign and how can you make more money like what were the good things and bad things about it so you integrate with the

company's social media platforms right there or you just scrape it for them i scrape it for them if they want to integrate it can be integrated especially the founder -led brands um i want integration i haven't been able to onboard a founder -led brand yet i want integration so that

i can really see like okay today you did a post and it was controversial and your your sweatshirt sales went down i really want to see the data behind that especially like i said in this new creator economy that's going to be very interesting to have so this is the most

Roadmap, Costs, and Open Questions

fun part of the I think of the marketing operating system that I've built but it's an intelligent campaign so you basically just put in a product and does everything for you but this part is a really interesting moat if I will so no agentic behind it yet you have to tell it to do stuff for you as opposed to

for it to take action by itself yes you have to tell it to do things for you there is I can just flip a switch and it's a ginseng but right now I'm telling it to do things until I figure out how to best direct it to do things for me

thank you so this yeah this is because you see the local host this is just demo data but I actually have it in super base right now where I'm hosting it I

just finished an incubator accelerator program with UM and then I'm into other incubator programs well I will pitch to hopefully win 100k and put more money behind it so I can increase the database and really start getting the AI to learn

more yeah it seems like it's really good use case for a second brain concept that you explored that are you familiar no I would love to talk to you about that

yeah yeah yeah for sure any other questions no this is so fun it's really something that of course if you want me to do it but it's really something that you can easily adopt you can go back and say I watched this girl have this presentation and she did XYZ and cloud can really take you step by step and

and create something really local for you to get this done for cheap.

Tooling Costs and Staying Efficient

I pay $20 max. I think my Higgs field subscription is 40, but overall my subscriptions are less than $200

between Cursor and Higgs field and all the things. So I really don't spend that much to get a lot of output.

Agents, Local Models, and Digital Twins

Are you leveraging Ollama at all for local models? Am I what? Leveraging Ollama and local models? No.

You could drive the cost down probably even more. No. Okay, yeah, we definitely have to talk. I would love to drive my costs down.

I need more agents. I need to put more money behind agents.

Do you use any agentic harnesses like open claw or herpes? no I yeah no I'm doing everything manual now um I was just advised to learn so much about it that when I start paying and implementing it that it has a really good direction which is wasn't bad

bad advice any questions cloning like like a digital twin yeah yeah so I meant to go over that thank you for asking that question I wanted to do a demonstration today but we don't we are short on time so it's very easy to do it

take a picture of all angles upload it to Higgs field create a character and and then the character sheet that I showed you earlier, and then you can insert an identical digital twin into any space and it looks real.

Whoa. Yeah, it's very interesting.

If you go on Higgs Phil's website or YouTube, the guy who pops up, he's not real. Like he's real, but majority, a lot of clips in there are his digital twin and it's crazy.

Conclusion

Any other questions?

All right. All right. Thank you guys so much.

Finished reading?