DEV Community: Seb

This is what every AI pin gets WRONG

Seb — Sat, 29 Aug 2026 14:41:36 +0000

If you are interested to get this project developed further, cosider starring it on my github: life-autopilot-agentic

WITH the way things stand, AI pins promised the FUTURE, but the technology simply couldn’t keep up with the demands and expectations. We expected these devices to be as responsive as our phones and watches. Yet many of them became little more dead weight, so much so, that you’re better off just hanging it on the wall as a moment in history.

It’s easy to speak to someone on loudspeaker or through an earbud because we get instant response. These AI pins promised the same convenience and more. but to speak to an AI model that takes longer than 3 seconds to response, quickly becomes uncomfortable for others around you, and probably you too. Even the addition of screens and laser projectors, while visually impressive, did little to improve the actual usefulness of the products. The vision was exciting, but with the combination of systems contained in this tiny device with a tiny battery, it was bound to be an immense challenge. for one, overheating was a problem, which personally is a shocker, considering we have smartwatches. Granted this ‘AI pin’ industry is very new, so these kinks are inevitable. But do those kinks mean that there is no place for ai pins in this era?

If They Failed So Badly, Why Did Everyone Want One?

Firstly, based on the response received when these products were announced, is a clear indication that the desire to use these devices are there, not only were investors invested, but also consumers. Secondly, considering that we have strong performing smartwatches loaded with various sensors and devices, it would not be hard to recreate a product of a similar form factor. Smartwatches have screens, they are packed with sensors, they have mics and speakers, they can use wifi and data, and they have all day battery life. so these watches are more than capable of functioning like an ai pin, so what stopping anyone from making a specialized ai pin, you might think. It would be nice to have this reliable ai system that is always aware of what’s happening around you and ready to help you out as we were promised.

The Question That Started This Deep Dive

My goal is to understand what resulted in these products performing so poorly despite generating so much excitement. To do that, we'll look at the setup of each product, compare smartwatches and AI pins, and identify where things went wrong. These deep-dives will reveal if there was a better way to execute these products, or if the large response seen by the announcement of these product was just a misleading sign to the demand of this innovative product. The comparison will be between the infamous Humane AI pin and the Apple smartwatch, because it seems fitting that the founders of Humane were former Apple workers.

Now these ‘deep-dives’ that I will do, will basically be surface level observations, because the most common reason for a products success or failure are things that have a direct impact on experience. So therefore, I will examine first the hardware, things like material and parts. After that I will examine the usability of each part of the product then the usability of the whole product. Then the last part I will look closer at is conveniency, how did the product strive to offer convenience and how was the overall outcome of its attempt to do so. Then in the end I will show and tell you about my own take on how ai pins should be done.

Humane Built the AI Pin Everyone Was Talking About

In comparison to the long history that smartwatches have, being that the modern smartwatch first appeared in 2013 by a kickstarter, ai pins are just getting started. they were only introduced in 2023, just three years ago. The Humane AI pin appeared to be like the apple of an innovative ai forward product. It look luxurious, it had no screen but a laser , it seemed to be very responsive according to the ads, and they were made by ex-Apple veterans. But I think that the fact that it was made by former ‘Apple’ employees, it was almost like a sign that the AI pin would flop, considering that Apple itself basically flopped hard in its AI development compared to its competitors. Humane built their AI pin to be

a smartwatch-sized device made of aluminum which attached magnetically to clothing,
with no screen, but with a laser-ink projector as the interface.
Obviously it had a small battery.
It had a wide-angle camera to constantly stream to the AI.
Some mics to chat to it.
4G sim connectivity and some modern dual-band wifi and bluetooth.
Also they packed a smartphone-grade processor into it.

The Red Flags Started Immediately

Now already off the mark there are too many red flags about this. Why use a smartphone-grade processor for such a tiny device we aren’t there yet. there are processors like the Snapdragon W processors, specially made for wearables so the battery does not tank after seconds of use. Also it dangling on a t-shirt would obviously become a bother, and the magnets would not be strong enough to hold on something like a thick jacket. This is where they let their creativity over crowd their intuition of common product experience.

A Companion That Never Stopped Watching

The ai pin had gesture tracking via the camera which again, was constantly streaming, so it was always active, burning power, just waiting for you to do a gesture, or ask a question about something you are showing it or what it saw earlier in the day. It was positioned to be a companion but it was not built to be one. As the camera is live, it is also querying the ai about what was observed, and this is done both on the device as well as across the internet. So the battery is taking a beating and the querying time is an eternity. A more obvious solution would be to pair it with a smartphone, but at that point you might as well use the phone, lol.

The Coolest Feature Was Also a Gimmick

The edition of the laser projector is another case of humane letting creativity flow uncontrollably with no consideration of usability. While the idea looked futuristic in demonstrations, the projector struggled in real-world lighting conditions. As a result, it became more of a gimmick than a practical interface.

The Smartphone Replacement Problem

Now the sim connectivity and bluetooth bundled into this device seemed to serve no purpose at all in this device. Smartphones are great, they are fundamental. They were built to handle the essential aspects from the beginning and they have been perfected to handle them, and they can handle many more functions.

So proposing a full blown replacement in the size of a ‘pin’ is unrealistic. We have smartwatches and even they have not replaced smartphones.

Just from this breakdown, the whole product is riddled with red flags.

Meanwhile, Apple Took a Completely Different Approach

Now switching to smartwatches, all the way back in 2015, Apple released their smartwatch, and they presented it as a luxury accessory, packed with some trackers.

They were built to be strapped to the wrist and they had bluetooth, a low energy edition of it, with the main purpose of syncing with you phone, wifi was only added some years later, for scenarios where you did not have you phone in range.
They were built with an led derived screen
They had a small battery, and today still do.
They also were built of aluminum.
They were packed with sensors and tech like accelerometers, gyroscopes, digital compasses and even heart rate sensors.
Apple made their own wearable chip, but some other manufacturers used mobile chips but later developed wearable chips too.

Why Smartwatches Actually Worked

As the years went by smartwatches just became more mainstream, thus more features were added and other things were improved in later generations. But we only need to look at the impression that the first version made.

On a small battery, the smartwatch was able to last closer to 20 hours that 10. They were really smart because of the gyroscopes and accelerometers built into them. This allowed unique features to be form by transforming the motion data they tracked. It had a smaller screen so you could do fewer things. Which helped distance us from using smartphones. since they were strapped to our wrists there was some way to dissipate heat. The first edition screens had poor efficiency, so it would really drain the battery over long usage. But that still allowed for hours of usage.

There really is little or even nothing to complain about smartwatches. They were so well built that it led to their unprecedented success. Unlike the AI pin which suffered with latency, efficiency, and redundancy.

The AI Pins That Learned the Right Lesson

But as you might know, humane is not the only AI pin product in the market. Most of the alternatives are screenless wearables that act as a side presence and not the latest thing to replace your smartphone. like

the Friend ai necklace. Built strictly for companionship rather than productivity, offering unprompted thoughts, commentary on your day, and emotional support as a persistent "friend".
The limitless pendent, Designed entirely around "life-logging" and professional note-taking

The reason they weren’t the main focus is because they were not talked about as much as the humane ai pin was.

The alternative AI pins really do execute the ideal way the AI pin should be. In the background and helpful.

So What Should an AI Pin Actually Be?

Now if you’ve made it this far, it’s time for me to explain what I think AI pins should be like. The interesting part is that my version doesn’t require any revolutionary hardware. So to get the point across, picture yourself as a recently employed worker at a well-paying company. It's the weekend and you're lying on the couch scrolling through social media. Life is going well. You've settled into your new job, your finances are looking healthier, and for the first time in a while it feels like you're building something for yourself.

While scrolling, you come across a local hiking group organizing a trip for next Saturday. The photos look incredible, the route isn't too far away, and on impulse you decide to join. You save the event and move on with your day.

That's it. You don’t open an app, or create a task, or set reminders, or even tell your assistant, "Hey, remind me about this next week."

What If Your AI Understood What You Were Trying to Do?

Now the AI that actually understands what you're trying to do.

Throughout the week, it quietly pieces together the situation. It knows you work Monday to Friday. It knows how long your commute usually takes. It notices that on weekends you tend to wake up later than you do during the week. It sees where the hiking event is and notices that rain is expected around midday.

Then Friday evening arrives.

Most AI assistants would sit there doing absolutely nothing until you ask a question. This AI doesn't. You're making dinner when it casually speaks through your earbuds.

"Hey, that hiking trip you're going on tomorrow looks like it's going to be fun."

You pause. "Oh yeah, I almost forgot about that."

"I figured. By the way, if you leave around seven fifteen tomorrow you'll get there comfortably before everyone else. Also... maybe pack something waterproof. The forecast changed this afternoon and it's looking like rain around lunch."

The Secret Isn't Faster AI

What it just did was a co-ordinated participation in your life. A few seconds later you ask, "How bad is the rain supposed to be?" It responds, “Enough that you'll probably be annoyed if you're standing on a mountain without a jacket."

You laugh.

"Fair point."

"There's a camping store on the way home from work. They're still open if you need one." What's interesting is that the assistant isn't improvising these answers one question at a time. The moment it decided to tell you about the hike, it had already prepared answers to the questions you were most likely to ask next. Can you see how different that feels. The agent doesn’t overload you with every detail upfront. Instead it gives you the most important info first then predicts what you can about. If you ask a follow up question, the answer appears instantly because it already did the thinking, its now just matching and returning the response.

A Real Assistant Knows When to Speak

The next morning, things get even more interesting. You oversleep. Your alarm goes off, but you hit snooze. Then again. And again. The AI waits for a response but notices that you are not giving a response. So a voice comes through your phone. “Wake up. You know if you keep pretending to be asleep, you're going to miss that hike. "you groan and ask what time is it?" it responds saying, ‘later than you wanted it to be." You sit up. "Oh no."

The Agent tells you still have time. Get dressed, skip the extra coffee stop, and you'll make it. As you're driving, traffic starts building because of roadworks. Before you even notice, the AI speaks again. "Bad news. The route you're on just got slower."

You ask how much slower, and it tells you and simultaneously offering a different route. That's the difference. Most AI products today act like search engines with personalities.

You ask, they answer, you command, they obey.

But that's not how a real assistant works. A real assistant pays attention, understands your objectives, it notices problems while they're still small.

Most importantly, a real assistant knows when to speak. The idea isn't an AI pin that lets me talk to ChatGPT without taking my phone out of my pocket. Because you will. The idea is an AI that understands enough about my life to casually lean over and say,

"Hey, if you leave in the next ten minutes, you'll make it."

Or even, "You've been saying you want to start exercising. The weather is perfect, your calendar is clear, and there's a walking trail five minutes away. i rsvp’d you for tomorrow’s run.”

Not because I asked.

Because it thinks it would help.

The Future of AI Companions

Finding ways it integrate AI systems into our life in meaningful ways is challenging. So it’s best to start with what is already available, and make improvements as we go. I have already done some development on this idea which you can try on your own phone by following the step on my GitHub. AI pins are still being experimented with, different form factors, different use cases, because people are trying to differentiate themselves. But the moment when a set of features has been proven to be necessary, then a complete, well-functioning product can be built, which can then be built in different, ways. Much like there are smartphones from different companies built to be unique but still contain all of the essential features.

Building Platter Pantry: MeDo Hachathon Project

Seb — Thu, 14 May 2026 08:53:33 +0000

Try The Demo With Your Kids, No Account Needed!

https://app-bd0ow6iaa9dt.appmedo.com

Most homes already have a pantry problem.

Food expires quietly in the fridge. Groceries get bought twice because nobody remembers what’s already at home. Ingredients disappear without anyone noticing. Shopping trips become incomplete because people forget what’s running low.

And despite all the “smart kitchen” ideas out there, most inventory apps fail for one simple reason:

Logging food manually is boring.

That was the starting point for Platter Pantry.

The Scale of Household Food Waste

The more I researched food waste, the more surprising the numbers became.

According to the United Nations Environment Programme (UNEP), households are responsible for the majority of global food waste. The UN estimated that in 2022:

1.05 billion tonnes of food were wasted globally
60% of that waste came from households
and the average person wastes 79kg of food every year. (Stop Food Loss and Waste)

The economic impact is massive too.

A 2024 UN report estimated global food waste costs around:

$1 trillion annually. (The Guardian)

Some household-level estimates are even more eye-opening:

U.S. households reportedly lose around $1,800 per year to wasted food
UK households waste around £700 yearly
and food waste globally may cost households over $1 trillion collectively every year. (Worldmetrics)

What stood out to me was this:

Most food waste doesn’t happen in factories.
It happens quietly inside homes.

And a lot of it comes down to visibility.

People forget:

what they bought,
what’s expiring,
what’s already in the kitchen,
and what ingredients they actually use.

That became the core problem Platter Pantry tries to solve.

The Original Idea

Platter Pantry started as an attempt to answer a simple question:

“What if households could actually understand what food they have in real time?”

The idea was to create a system where users could:

track groceries,
log meals,
monitor pantry levels,
reduce waste,
and get smarter shopping suggestions.

The app uses AI to parse:

receipts,
text,
voice,
images,
and even video(still in progress)

to automatically detect food products, quantities, and expiry estimates.

Instead of manually typing:

“2kg rice, 1L milk, 6 eggs”

you could simply upload a grocery photo or receipt and let the AI structure the pantry automatically.

The Problem We Ran Into

Even with OCR, speech-to-text, and multimodal AI integrations, there was still a problem:

Somebody still has to do the logging.

And in a real household, that responsibility usually falls onto one person.

That creates friction.

The more friction there is, the faster people stop using the app.

So instead of asking:

“How do we make logging easier?”

I started asking:

“How do we make logging collaborative?”

That changed the direction of the entire project.

Introducing Resident Loggers

This led to one of the biggest features in the app:

Resident Loggers.

A Resident Logger is a household member who helps manage pantry logging tasks.

Instead of one parent manually tracking everything alone, family members can now:

accept pantry tasks,
scan groceries,
upload receipts,
log kitchen items,
and submit pantry updates for approval.

The household owner reviews the submission, approves it, and the pantry updates automatically.

That transformed the app from:

“a pantry tracker”

into:

“a collaborative household system.”

Making Household Contribution Feel Rewarding

The more I explored this idea, the more I realized something important:

Kids already help around the house.

But most household chores feel disconnected from:

technology,
progress systems,
incentives,
or visible contribution.

So instead of designing the Resident Logger system like a boring admin panel, I started treating it more like:

a progression system,
a lightweight task platform,
and a gamified contribution tracker.

Resident Loggers now have:

XP and levels,
streaks,
badges,
accuracy scores,
task history,
earnings tracking,
and household contribution stats.

The goal is to make helping at home feel:

rewarding instead of repetitive.

Kids Can Earn Real Money Through Household Contribution

One of the most interesting parts of the system is that Resident Loggers can optionally earn real money for completed tasks.

For example:

logging monthly groceries,
scanning receipts,
organizing pantry inventory,
or helping update kitchen stock

can become paid household tasks.

Parents can:

assign rewards,
track completed work,
approve submissions,
and release payments.

This creates something I find really interesting:

household contribution becomes visible, trackable, and rewardable.

Instead of random chores, children participate in:

real systems,
operational workflows,
accuracy-based tasks,
and collaborative household management.

The app turns pantry management into something closer to:

“earning through contribution”

rather than:

“being told to do chores.”

Why I Think This Matters

Most productivity software is built for offices.

Very little software is designed around:

family collaboration,
household operations,
or teaching responsibility through real systems.

Platter Pantry sits somewhere between:

AI automation,
household coordination,
and gamified contribution systems.

And I think there’s something powerful there.

Not because pantry tracking itself is exciting,
but because:

homes run on invisible systems.

And most of those systems are still unmanaged.

Potential Problems This Could Solve

🥦 Food Waste

People forget what they already have.

The app tracks:

inventory,
expiry dates,
pantry levels,
and ingredient usage.

🛒 Smarter Shopping

Users can understand:

what runs out often,
what gets wasted,
and what actually needs restocking.

👨‍👩‍👧 Shared Responsibility

One person no longer has to manage the kitchen alone.

Families collaborate inside the same system.

🧒 Teaching Kids Responsibility Through Real Systems

Children don’t just complete arbitrary chores.

They contribute to:

real household operations,
measurable workflows,
accuracy-based systems,
and task completion pipelines.

🤖 Making AI Useful in Daily Life

This project focuses on something extremely practical:

helping households organize food.

The Current Direction

Right now, the platform is focused on:

Family-first collaboration.

Resident Loggers currently operate only inside households for simplicity and safety.

No public neighborhood marketplace yet.

Long-term, I’m interested in exploring:

trusted neighborhood support systems,
collaborative pantry management,
and hyper-local household logistics.

Final Thought

I didn’t start building Platter Pantry because pantry apps are exciting.

I started building it because:

homeowners buy more than they need which leads to waste of food, and more importantly, money.

There just are not many really good tools to manage this yet.

And if AI is going to become part of everyday life, I think one of the most meaningful places it can start is the home.

This AI Is a Better Geography Teacher Than Most Humans.

Seb — Sat, 13 Sep 2025 15:45:51 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

I built GeoGenius, an AI-powered interactive learning companion designed to make learning geography dynamic, engaging, and fun. The problem it solves is simple: traditional learning from static textbooks can be dry and fails to capture the dynamic nature of our planet. GeoGenius transforms any geography topic—from a simple text prompt or even a photo of a textbook page—into a personalized, multi-faceted learning module in seconds.

The user flow is straightforward:

A user enters a topic like "Causes of Urbanization" or uploads their study notes.
They click "Generate Lesson."
GeoGenius uses AI to create a rich, structured lesson plan that includes:
Key Concepts: A clear, concise breakdown of the most important ideas.
Simple Analogies: Complex topics are related to everyday life to make them easier to understand.
Visual Moments: Each analogy is paired with a unique, AI-generated image to visually reinforce the concept.
Interactive Animations: For dynamic processes like the water cycle or tectonic plate movement, a simple, hands-on animation is provided.
Test Your Knowledge: A short, multiple-choice quiz helps the user check their understanding and solidify what they've learned.

It's not just a Q&A bot; it's a creative partner that builds a personalized and comprehensive lesson from scratch, every time.

Demo

link to the applet: https://geogenius-interactive-geography-lessons-598974168521.us-west1.run.app

For a complete walkthrough, here is a short video that shows the entire process from entering a prompt to interacting with the final generated lesson.

How I Used Google AI Studio

Google AI Studio was the core in the development of GeoGenius, particularly for prototyping and refining the core AI prompts.

My workflow was centered around structured output. I knew I needed a reliable JSON object from the AI to render the different components in my React app.

I used the AI Studio playground to:

Craft the System Prompt: I iterated on the main instruction ("You are a world-class geography teacher...") in AI Studio to find the perfect persona and set of instructions for the model. This ensured the tone and quality of the content were consistently high.
Define and Test the JSON Schema: The "Structured Prompt" feature was a game-changer. I designed my entire lessonPlanSchema directly in AI Studio, defining the types, required fields, and even providing descriptions for each property. This allowed me to test prompts and see how the gemini-2.5-flash model would populate the schema, ensuring the output was always predictable and ready for my frontend to consume.
Prototype Multimodal Inputs: I also used the playground to test how the model would respond when given an image (like a page from a textbook) along with a text prompt, which helped me build the file upload feature with confidence.
After perfecting the prompts and schema in AI Studio, I simply transferred that logic into my application code using the @google/genai SDK. This process saved hours of development time by allowing me to separate prompt engineering from application logic.

Multimodal Features

GeoGenius leverages multimodality to create a richer and more effective learning experience than what text-alone applications can offer.

Multimodal Input (Image + Text → Lesson): The app’s ability to accept an image (like a .png or .jpg of a map or textbook page) alongside a text prompt is its most powerful multimodal feature. A student can literally take a picture of their homework and ask the app to build a lesson around it. This turns static, offline content into a dynamic and interactive digital experience, meeting the user exactly where they are.

Text-to-Structured Data (JSON): At its core, the app transforms an unstructured text request into a highly structured JSON object. This is a crucial modality that allows for a reliable and sophisticated UI. Instead of getting a single block of markdown, the AI provides a predictable data structure that the React frontend can map to distinct, beautifully styled components for concepts, analogies, quizzes, and more.

Text-to-Image Generation: To make abstract concepts more tangible, the AI first generates a creative visualPrompt for each analogy in the lesson plan. My application then takes this text prompt and uses the imagen-4.0-generate-001 model to generate a custom illustration. This on-the-fly image generation ensures that every lesson has unique, contextually relevant visuals, which are far more engaging and effective for learning than generic stock photos.

Can an Algorithm Dream? I Built an App to Find Out.

Seb — Tue, 09 Sep 2025 04:34:53 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

I built the Progressive Story Maker, an interactive web application that transforms storytelling into a collaborative, choice-driven experience with an AI.

The app solves the problem of "writer's block" and creative inertia by turning narrative creation into an engaging game. It begins by generating the first sentence of a story in a user-selected genre (Medieval Fantasy, Modern Mystery, or Kiddish Adventure). Within this sentence, key words are highlighted. When the user clicks a word, it becomes the creative prompt for the Gemini API, which then generates the next paragraph of the story. This new paragraph has its own clickable keywords, allowing the user to continuously guide the narrative down unique, branching paths. The result is an endless story machine that empowers users to co-create completely original tales simply by making a series of simple, intuitive choices.

Demo

here is the link: https://progressive-story-maker-598974168521.us-west1.run.app

How I Used Google AI Studio

While the final application interacts directly with the Gemini API via its SDK, Google AI Studio was an indispensable tool during the development and prototyping phases.

Prompt Engineering & Refinement: I used the AI Studio playground extensively to design and test the prompts that power the application. It provided a rapid feedback loop for crafting instructions that could reliably generate compelling story segments and, most importantly, extract exact, verbatim keywords from the generated text. This was crucial for ensuring the frontend could always find and highlight the interactive words.

Structured Output (JSON Mode): AI Studio was instrumental in defining and validating the responseSchema for the Gemini API calls. By experimenting in the studio, I finalized a robust JSON structure ({ "paragraph": "...", "keywords": [...] }). This use of structured output makes the application incredibly resilient by guaranteeing that the data received from the API is always predictable and correctly formatted, eliminating the need for fragile string parsing and significantly reducing potential runtime errors.

Model Selection: I used AI Studio to evaluate different models, ultimately selecting gemini-2.5-flash for its optimal balance of speed, creativity, and cost-effectiveness, which is essential for a real-time, interactive user experience like this one.

Multimodal Features

The current version of the Progressive Story Maker focuses on perfecting a masterful unimodal (text-to-text) experience to ensure the core narrative mechanic is seamless and engaging.

However, the application was architected as a strong foundation for future multimodal expansion.

Image-Driven Prompts (Image-to-Text): To introduce multimodal input, users can upload an image instead of clicking a keyword. The contents of the prompt sent to Gemini would then include both the image and a text instruction like, "Continue the story based on this image." This allows users to introduce completely new visual concepts into the narrative, giving them an even more powerful way to guide the AI's creativity and making the experience truly multimodal.

The Ultimate Life Hack for Turning a Boring Walk into a Real-Life Video Game Hometown

Seb — Tue, 09 Sep 2025 03:36:44 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

I used Google AI Studio to make City Quest AI is an interactive, AI-powered scavenger hunt application that transforms any city or neighborhood into a dynamic adventure. The app solves the problem of generic, pre-planned city tours by creating unique, on-the-fly quests based on user input. Users simply enter a location (e.g., "Downtown Seattle") and an optional theme (e.g., "Historic Landmarks" or "Coffee Shops"), and the app generates a multi-stop scavenger hunt complete with clever clues, fun challenges, and interesting facts for each location.
It's designed to encourage local exploration, create fun and competitive experiences for groups of friends, and make discovering a new place more engaging and memorable.

Demo

a link to my deployed applet: https://city-quest-ai-598974168521.us-west1.run.app

Setup Screen: The user is greeted with a sleek, modern UI. They enter their desired city and an optional theme for their adventure.

Loading Screen: After submitting, a loading animation appears while the Gemini API crafts a custom scavenger hunt in the background.
Game Screen: The main interface displays the first clue. Once solved, the user can reveal the location, which appears on an embedded map. They then complete a fun, interactive challenge (like taking a photo) and learn a fun fact about the place. The screen also features a live scoreboard, adding a competitive element to the game.

Completion: After finishing all the stops, the user sees a congratulations screen celebrating their victory and encouraging them to play again.

How I Used Google AI Studio

I leveraged the Gemini 2.5 Flash model through the @google/genai SDK to serve as the creative "game designer" for the app. The core of the integration lies in using structured output capabilities.

In the geminiService.ts file, I defined a strict JSON schema (responseSchema) that outlines the exact data structure needed for each scavenger hunt stop: clue, locationName, challenge, and funFact.

By providing this schema in the API call along with a detailed prompt, I instruct the Gemini model to return its creative output as a perfectly formatted JSON array. This is powerful because it eliminates the need for unreliable string parsing and ensures that the data from the AI can be directly and safely loaded into the application's state, making the integration seamless and robust.

Multimodal Features

The primary AI feature is Text Generation with Structured Output. The application takes simple text input from the user (location and theme) and uses the Gemini model to generate a complex, structured JSON object that defines the entire game. This goes beyond simple text generation by forcing the AI to think and create within a specific, developer-defined framework.

While the AI generation is text-based, the application creates a multimodal user experience. The AI-generated challenge text often prompts the user to interact with the physical world using their device's camera (e.g., "Take a team photo in front of the fountain"). The user then uses the built-in photo upload component, which accesses the device's camera, to provide visual proof of completion.
This creates a fun loop:

Text-to-Game: AI generates a text-based challenge.
Real-world to Image: The user interacts with the real world and captures it as an image.
Game Progression: The image upload confirms the challenge is complete and unlocks the next part of the adventure.

Give it a try!

Seb — Tue, 09 Sep 2025 03:27:15 +0000

Google AI Challenge Submission

Seb

Sep 6 '25

Don't Wait for an Accident. This AI Tool Spots Hazards Before Your Baby Does.

#devchallenge #googleaichallenge #ai #gemini

3 min read

Give it a try

Seb — Tue, 09 Sep 2025 03:25:06 +0000

Google AI Challenge Submission

Seb

Sep 7 '25

This AI Guesses Your Drawings Faster Than Your Friends Can.

#devchallenge #googleaichallenge #ai #gemini

2 min read

This AI Guesses Your Drawings Faster Than Your Friends Can.

Seb — Sun, 07 Sep 2025 16:21:06 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

I built an interactive web application that brings the classic drawing and guessing game into the digital age with a modern twist. The app challenges a player to draw a word provided by the game, while a sophisticated generative AI attempts to guess the drawing in near real-time.
This creates a unique and engaging solo-player experience where the user's artistic skills are pitted against the AI's image recognition capabilities. It solves the problem of needing multiple players for a game of Pictionary and provides a fun, interactive way to experience the power of multimodal AI.

Demo

link: https://ai-pictionary-598974168521.us-west1.run.app

How I Used Google AI Studio

I leveraged the Gemini API, accessible through the @google/genai SDK, to power the core guessing mechanic of the game. Specifically, I used the gemini-2.5-flash model for its speed and powerful multimodal capabilities.

The implementation involves capturing the user's drawing from the HTML canvas as a PNG image, converting it to a base64 string, and sending it to the Gemini model. This image is sent alongside a carefully crafted text prompt: "What is this a drawing of? Look at the image carefully and provide your best guess in a single word." The model then processes this combined visual and textual input to return its guess as a single word of text. This demonstrates a powerful image-to-text, or visual understanding, use case.

Multimodal Features

The central multimodal feature of this application is visual reasoning and description. The app seamlessly integrates two distinct modalities:

Image Input: The user's free-form drawing on the canvas serves as the primary visual input.

Text Output: The Gemini model analyzes this visual information and generates a textual guess.

Don't Wait for an Accident. This AI Tool Spots Hazards Before Your Baby Does.

Seb — Sat, 06 Sep 2025 05:03:08 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

BabySafe AI is a user-friendly web applet designed to give parents and caregivers peace of mind. It solves the often overwhelming problem of childproofing a home for a mobile baby (6-18 months) by acting as an AI-powered "second set of eyes."
Users simply upload a photo of a room, and the applet leverages the Google Gemini model's vision capabilities to analyze the image, identify common household safety hazards, and present them in a clear, actionable list. The experience is designed to be simple, fast, and reassuring, turning a potentially stressful task into a straightforward safety check.

Demo

link: https://babysafe-ai-598974168521.us-west1.run.app

Here's a walkthrough of the BabySafe AI user experience:
Upload: The user is greeted with a clean, simple interface where they can drag-and-drop or click to upload a photo of a room.
Preview & Analyze: Once an image is selected, a preview is shown. The user then clicks the "Analyze Safety" button to begin the AI analysis.
Analysis in Progress: A clear loading indicator lets the user know the AI is working.
Hazard Report: If hazards are found, the results are displayed in a two-panel layout. The original image is on the left, and a detailed, scrollable list of identified hazards is on the right. Each item in the list includes the hazard's name, the specific risk it poses, and a clear description of its location.

How I Used Google AI Studio

I used the gemini-2.5-flash model from the Google AI platform to power the core functionality of BabySafe AI. My implementation focused on two key multimodal capabilities:
Vision Understanding: The primary function relies on the model's ability to interpret and understand the contents of an image. It goes beyond simple object detection to recognize items in the context of child safety.

Structured JSON Output: I used Gemini to write up a detailed system prompt and provided a strict responseSchema. This instructs the AI to act as a "Baby Safety Expert" and forces it to return its findings in a predictable JSON format. This was crucial for reliably parsing the AI's response and rendering the hazard list in the user interface. By defining the expected data structure, I turned a powerful, general-purpose vision model into a specialized and reliable analysis tool.

Multimodal Features

The core multimodal feature of this applet is Image-to-Structured-Text Analysis. It takes a visual, unstructured input (a user's photo) and transforms it into structured, actionable text data (a JSON object listing hazards).

This enhances the user experience in several key ways:
Intuitive Input: Users don't need to describe their room with words. They can simply show the AI the environment with a photo, which is a far more natural and efficient way to communicate complex visual information.

Contextual Hazard Identification: The AI doesn't just list objects; it identifies them as hazards. It understands that an electrical outlet low on a wall is a risk to a crawling baby, or that a dangling cord is an entanglement threat. This contextual understanding is only possible through analyzing the visual data.

Spatially-Aware Descriptions: The model generates a location_description for each hazard (e.g., "On the floor in the bottom-left corner"). This spatial awareness, derived directly from the image, makes the feedback immediately actionable. The user knows exactly where to look and what to fix, transforming a generic warning into a specific, helpful instruction.

A social media app, powered by BrightData & n8n

Seb — Tue, 26 Aug 2025 10:44:11 +0000

Ok this project is built to eventually be submitted to the Real-Time AI Agent Challenge.

This web application pulls news headlines, it is given to an LLM that is tweaked to exhibit different view points. After this it gets sent as a post on the app for everyone to see.

No one can make a post, but they can comment and show if they agree with the opinion of the model.

Currently this app is like 80 percent there. i just need to setup authentication and the design.

I really want to see this app atleast get 100 people on it. So please let others know.

AI Helped Me Build a Redis-Powered IoT Anomaly Detector. In 1 Day.

Seb — Sat, 02 Aug 2025 01:42:50 +0000

This is a submission for the Redis AI Challenge: Real-Time AI Innovators.

What I Built

While vide-coding, I made a Redis-Powered IoT Anomaly Detector in one day." It demonstrates a real-time data processing pipeline designed for an Internet of Things (IoT) sensor, using Redis Streams and Redis Time Series. The system ingests continuous sensor data streams, detects anomalies through a threshold-based logic, stores data efficiently for historical analysis, and raises alerts for critical events in real time.

Demo

https://github.com/SebastianKibiriti/redis-iot-anomaly-detector-mvp-withGemini

How I Used Redis 8

It uses Redis Streams for real-time data ingestion and inter-service messaging, and Redis Time Series for optimized, long-term storage and analysis. Two key Redis features that are central to Redis 8's capabilities for IoT and time-series applications.

What are Your Daily Driver Apps??

Seb — Thu, 24 Jul 2025 16:00:04 +0000

So Gemini has a feature to use the search history of the user for context. I never thought how I could use it, until now. This post has given me a reason to try out that feature. Considering the last 30 days, I've been pretty using the same apps/software for developing applications. Despite the advances that the competitors make, the tools I use are pretty much comparable to the competition and I don't notice a performance difference. I just use them since I am used to them.

Tools I used in the last 30 days.

Large Language Models/AI Assistants:

Gemini
ChatGPT

AI Image Generators:

Ideogram
Higgsfield

AI Terminal:

Gemini CLI

AI Video Generators:

Higgsfield

AI Development Platform:

Bolt.new
VS Code
Google AI Studio