David Castro

Founded an AI study platform that turns notes, slides, and documents into quizzes, mock exams, and study notes. Started in November 2023 on OpenAI's text-davinci-003, a completion model that predated the chat-based LLMs everyone uses now. The model couldn't handle long documents, so I independently figured out a way to split content into chunks and process them individually, which turned out to be the same technique the industry later standardized as chunking. Over time I taught myself Firebase for authentication, Stripe for billing infrastructure, Railway for deployment, and Google Analytics for tracking. I also implemented PDF downloads, rate limiting, bot prevention, and SEO from scratch. Every major technical skill I use today traces back to something I had to figure out for CramIt. Currently freemium with 2,200+ learners served and 35+ daily active users.

Organized and hosted a multi-university AI ideathon at KMC, The Podium West Tower in Manila. I chaired the event end-to-end, from screening 40 applications across 15 Philippine schools down to six finalist teams, to running mentoring sessions and flash talks on using AI to accelerate idea development. The judging panel included Dominic Ligot of Cirrolytix and Shad de la Cruz of Symph. Winning projects ranged from an augmented reality game that teaches AI concepts to an AI-powered agricultural drone system. The event was co-presented by DashoContent and Web3PH and co-organized with Young Founders Summit, GEN AI Philippines, Third Team Media, and DashoWork.

Founded and built a geospatial intelligence platform that tells small Philippine internet service providers exactly which neighborhoods are worth deploying to. The problem I kept running into was that all the data you'd need to make that decision exists, but it's scattered across nine different government and scientific sources that were never designed to talk to each other. So I unified them. I pulled nighttime light data from NASA VIIRS, cell tower locations from NTC and OpenCelliD, poverty estimates from PSA, and population density from a JRC/Copernicus dataset that I got directly from a climate consultant at the EU Joint Research Centre after meeting him at a university event. I wrote 21 Python scripts to clean, spatially join, and score all 42,048 Philippine barangays across nine dimensions into a single ranked dataset. The entire frontend is a single HTML file with zero frameworks and zero backend, running MapLibre GL JS on a 199MB vector tileset served through Cloudflare R2 with HTTP range requests so the browser only downloads what's visible on screen. It also has a live Starlink satellite overlay that fetches orbital data from CelesTrak and runs SGP4 propagation client-side to show real-time satellite positions over the Philippines. This live demo of a report based on the data I have compiled covers Iligan City.

I was already finding ways to use text-davinci-003 for SEO and brand management before ChatGPT went viral. One of my proudest results was a brand voice scanner I made from scratch. Paste a link, and it would scrape the site, crunch the content, and extract the exact semantics and tone of a brand. I also brought our Draft Generator to our internal teams, getting our workforce comfortable with AI-assisted writing before ChatGPT even went viral. I used Streamlit for fast iteration so we could move quickly, and by the time the rest of the industry caught up, our team had already been working with these tools for months.

I led a three-person NASA Space Apps team to solve a challenge from NASA's Astrophysics Division: most exoplanets have been identified manually, and they wanted teams to create AI models that could do it automatically using NASA's open-source datasets. Our approach was to skip expensive custom model training entirely. Instead, the system searches through 17,589 real NASA planet records across the Kepler, K2, and TESS missions, finds the 25 most similar cases to any new signal, and uses a cheap language model to classify it based on those examples. It got 95 to 98% accuracy, beating a previous academic benchmark of 83% by over 12 points. Every result shows exactly which reference cases the AI looked at, so researchers can verify the reasoning instead of trusting a black box. Next.js frontend with a Flask ML backend.

A speed and combat-focused platformer I made in two weeks as a personal game jam. The core mechanic is time-stop: you freeze time mid-fight, dash through enemies, parry hitscan bullets back at shooters, and chain kills into combos. Levels are designed to be cleared in ten seconds. It has a recording system so you can replay your runs, a level editor, and randomly generated levels. I made it in p5.js, the modern successor to ProcessingJS. I grew up coding on Khan Academy's project gallery, which ran on ProcessingJS, and that community is where I first started making games. CODE WRAITH was my way of pushing myself to see how much I could ship in two weeks with everything I've learned since then. I'm planning to go back to some of my old Khan Academy programs, rebuild them from scratch, and start releasing them on itch.io.