This is a submission for the Built with Google Gemini: Writing Challenge
What I Built with Google Gemini
As an Information Technology university student from Piñas, Ecuador, my true passion lies far beyond writing software on a screen: I am absolutely fascinated by physical hardware. I love getting my hands dirty with jumper wires, microcontrollers, and bringing the physical and digital worlds together.
Throughout my university journey, I haven't just used Google Gemini as an autocomplete tool; I've used it as my Senior DevOps, Electronics Engineer, and Full-Stack Architect. To demonstrate this, I am showcasing two of my most complex engineering projects—one rooted in raw hardware, the other in complex software algorithms—and how I plan to merge them.
Project 1: The Hardware Hub – IoT BMI Diagnostic Station
I built a smart, low-cost, and open-source automated health kiosk that measures weight and height, calculates Body Mass Index (BMI), and logs a historical record for multiple users.
The architecture is highly integrated:
The Hardware Brain: I used an ESP32 development board. Using ESPHome, I configured it to serve a dual purpose. First, it acts as a Bluetooth Low Energy (BLE) proxy to actively scan and intercept weight data from a Xiaomi Mi Smart Scale 2. Second, it physically reads distance from a JSN-SR04M ultrasonic sensor mounted on the ceiling.
The Logic Engine: The ESP32 streams raw data locally to a Home Assistant OS server. I built custom Jinja2 template sensors to subtract the measured ultrasonic distance from the calibrated ceiling height to get the user's exact height. It then calculates the BMI and classifies the semantic result.
Cloud Storage & UI: The UI is a locked-down Kiosk tablet running a custom Home Assistant dashboard. Every measurement triggers an automated script that uses OAuth credentials to securely push a JSON payload directly into a Google Sheets database.
Project 2: The Software Brain – Yacuviña Sunset Predictor & Analytics
Beyond health tech, I built Algorithm Yacuviña 3.0, a highly specialized meteorological prediction system and analytics dashboard for a local archaeological viewpoint. It predicts whether visitors will experience a spectacular "Sea of Clouds" or a "Clear Panoramic" sunset.
The Core Algorithm: The backend concurrently fetches data from Open-Meteo, OpenWeather, and AccuWeather. It uses theSunCalclibrary to pinpoint the exact local sunset time.
The Dual-Scoring Model: It evaluates two distinct scenarios: the Sea of Clouds model (weighting low cloud density and visibility) and the Clear Sunset model (prioritizing total cloud cover and UV index).
The Frontend & Analytics: The client is a React/Vite app utilizing premium glassmorphism UI. The backend exposes JWT-protected analytics endpoints to track geo-distribution and unique API hits.
What Role Did Google Gemini Play?
For the BMI Station, Gemini 2.5 Pro was my virtual electrical engineer. In fact, I explicitly cited Gemini as my official "AI Assistant" in the project presentation document and technical paper submitted to my university! It mapped out my physical hardware, explained the ESP32 GPIO pinouts, and generated the precise text prompts I used to render Fritzing-style schematic diagrams. For Yacuviña, Gemini architected my Node.js backend. It helped me structure the JWT rotation logic, design the caching strategy to avoid rate-limiting the weather APIs, and optimize the CSS for the frontend.
Demo
Live Demo: Check out my deployed Yacuviña Sunset Predictor & Analytics Dashboard here!
(Note: Since my physical BMI setup is currently disassembled, here are the real screenshots and diagrams from my project documentation!)
The BMI Station in Action:
(Here is the physical IoT BMI kiosk scanning and calculating in real-time!)
What I Learned
This journey was a masterclass in the collision between perfect software logic and chaotic physical realities.
Technical Deep Dives: Connecting the ESP32 to the Xiaomi scale wasn't plug-and-play. I had to learn how to turn the ESP32 into a BLE "radar" (a Bluetooth scanner) to sniff out the exact MAC address of the scale amidst the wireless noise. For Yacuviña, I learned how to handle asynchronous data streams in Node.js using Promise.allSettled, ensuring my dashboard didn't crash if one of the weather APIs timed out.
Unexpected Lessons (The Magic Smoke): Software forgives; hardware absolutely does not. Gemini correctly taught me that the JSN-SR04M ultrasonic sensor outputs a 5V signal, which would instantly fry the ESP32's 3.3V GPIO14 pin. It taught me how to build a voltage divider circuit using two 1kΩ resistors to safely step down the voltage.
However, I was using old, cheap jumper wires that were internally broken. Frustrated by the lack of sensor readings in Home Assistant, I started blindly unplugging and plugging the wires directly into the breadboard. In my haste, I mixed up the positive and negative wires. Kaboom. A puff of magic smoke, and the 5V sensor was dead. It taught me that while AI can provide perfect pinouts, it cannot fix your faulty cables or save you from a reversed polarity mistake. Measure twice, plug once!
Google Gemini Feedback
Having Gemini 2.5 Pro (and now 3.1 Pro) by my side has been like having a Senior IoT and DevOps Engineer on call 24/7.
What Worked Flawlessly:
Gemini is an absolute beast when it comes to infrastructure syntax. Writing the script to authenticate Home Assistant with the Google Cloud OAuth API requires incredibly specific JSON/YAML structures. Gemini generated the exact formatting on the first try. In the Yacuviña project, it flawlessly generated the complex JavaScript logic to dynamically switch between the Sea of Clouds and Clear Sky algorithms based on cloud strata percentages.
Where I Ran Into Friction:
LLMs often assume that errors lie entirely within the code logic, completely missing the physical constraints of the real world. Before my ultrasonic sensor exploded, my ESP32 was returning null distance values. Gemini kept suggesting software fixes: "Increase the update_interval," "modify the median filters," "check the firmware buffer size." It took me a fried sensor to realize the code was perfect; my broken jumper wires were the culprit.
I had to learn a completely new way of prompting: "Assuming the YAML code is 100% correct, what physical hardware, breadboard wiring, or voltage issues could cause this specific sensor reading to fail?" Once I established strict physical boundaries in my prompts, Gemini transformed from a generic code generator into an outstanding hardware debugger.
Beyond These Projects: My Long-Term AI Co-Pilot
While these two projects are my proudest achievements, the truth is that Gemini has been my co-pilot throughout my entire university degree. Our history together goes way beyond Home Assistant and React.
Gemini was there when I had to configure a 4-node Big Data cluster using Apache Spark, Hadoop, and Jupyter. It helped me survive setting up an Oracle 19c database on Linux, configuring firewalld rules, proxies, and webmail servers. When I went down the rabbit hole of distro-hopping between Linux Mint, CachyOS, and Ubuntu, Gemini helped me rescue my Windows bootloader.
We’ve built point-of-sale systems ("pos-carmita-villegas"), tackled a road alert app for Ecuador deployed on Vercel ("Vialert EC"), and even developed "DevFlow," a custom CLI wrapper for GitHub Copilot CLI. From troubleshooting IPv6 dual-stack RIPng networks in Cisco Packet Tracer to diagnosing the sickest hardware bugs, Gemini hasn't just been an API I call for code snippets—it has been the mentor that taught me how to become a full-stack, DevOps, and IoT engineer.
Looking Forward: Bridging the Gap
My ultimate goal is to merge my hardware passion with my software algorithms. If I win the Raspberry Pi AI Kit in this challenge, my plan is to build a physical edge-computing weather station and deploy it directly at the Yacuviña archaeological viewpoint. Instead of pulling generic data from web APIs, I will wire real humidity, barometric, and temperature sensors to the Raspberry Pi. I will run the Algorithm 3.0 locally on the Pi, utilizing its edge AI capabilities to process hyper-local data in real-time.
Hardware is hard, and software is complex. But with a soldering iron in one hand, a keyboard in the other, and Google Gemini guiding the architecture, there is no limit to what we can build.
SOCIAL SHARE CARD GENERATOR