Gabriel Busto
things i’ve built
most of these were full end-to-end apps that were deployed, had a signup flow, and database. some were just prototypes or for fun.
- blocksmith ai (generate editable 3d block models from a prompt or image)
- research and learning workflows (research workspace and ai tutor)
- additional product experiments (video, creative analysis, and landing pages)
- nft art marketplace (built a full nft art marketplace as a startup)
blocksmith ai
model generation case study · open-source core · texturing writeup
blocksmith was a full-stack ai product for generating and texturing editable minecraft-style 3d assets. it began after i encountered an asset-creation bottleneck while building games for hytopia.
the first prototype converted images into voxelized meshes, but developer feedback showed that the output representation was wrong. game developers needed editable, low-complexity models composed of semantically meaningful cuboid parts. i rebuilt the system around llm-generated semantic entities and later introduced a pythonic dsl when raw json limited model complexity and cost efficiency.
the dsl compiled models into glb and blockbench-compatible formats. a separate multimodal texturing pipeline used mv-adapter on modal and custom back-projection to paint the model’s texture atlas with a pixelated style.
the product gained free and paid users, including game-development teams that created hundreds of assets. i later sunset the closed saas after usage, pricing, training-data constraints, and the investment required for the next quality leap no longer justified continuing it as a company. the core generation engine is now open source.
the project started with no prior 3d background and eventually required custom model generation, format conversion, diffusion, data, deployment, and animation systems. the linked writeups go much deeper into the experiments and dead ends.
research and learning workflows
ai tutor (duolingo-like web app) - offline
the ai tutor was an attempt to turn open-ended learning goals into structured, small-format lessons rather than asking users to navigate research material directly. it focused on helping people learn about ai in byte-sized chunks (yes, pun intended), but the underlying workflow could support other subjects too.
after a short onboarding about motivation, learning style, and tutor personality, the user entered a topic. the app researched it, proposed high-level modules, and generated short lessons with citations.
each module ended with generated fill-in-the-blank, multiple-choice, or short-answer questions. the app graded responses and provided feedback when the learner got something wrong.
research paper assistant (prototype)
this was a prototype that i never launched publicly or posted about, but it led to the ai tutor app above. it came pre-loaded with ai research papers, and you could open them up onto an infinite canvas. you could move and re-arrange the pages however you wanted.
on any given pdf page, you could click and drag to create a highlight box, then right click to either create a sticky note for that region, or start a chat with ai about that region. it would clip the image behind the highlight box, and send it to an llm along with context about what paper you’re reading and what page you’re on.
each paper that was “pre-loaded” was processed using unstructured.io’s open source chunking code, and weaviate as a rag database to store the chunks. i used litellm on the server as the tool using assistant/agent, and it could:
- search the internet using google’s search api
- query the weaviate rag database using several different methods (search across papers, search through a single paper, retrieve page contents, etc)
i wasn’t confident that the paper workflow represented a sufficiently urgent problem for a broad audience, so i reused parts of the work in the ai tutor above. that product focused on generated learning paths, short lessons, citations, adaptive questions, and feedback.
additional product experiments
-
video understanding - processed video into aligned frames and assemblyai transcripts, then supported grounded questions over the resulting timeline.
-
creative analysis - evaluated the relationship among a youtube title, thumbnail, and opening content to generate structured scores and feedback.
-
landing-page analysis - rendered, screenshotted, and parsed live pages, then generated recommendations around clarity and conversion.
an nft art marketplace (full web app + android tv app + custom smart contracts)
early in the nft market, my cousin and i built tabu art to make digital-art ownership more meaningful than a token pointing to a public image. he brought relationships with artists, and i left raytheon to build the product.
i entered the project without production full-stack, aws, or blockchain experience. over roughly a year, i built a django/aws marketplace with stripe payments, ethereum contracts, asynchronous workers, media processing, and an android tv app.
our product thesis combined protected high-resolution originals, credit-card purchasing, buyer-funded minting, royalty routing, and a display experience for collectors. the payment flow hid most blockchain complexity, estimated transaction costs, and refunded unused fees. a custom erc-1155 contract let buyers fund minting so artists did not need to pay upfront gas costs.
the media system created device-appropriate image derivatives, encrypted purchased artwork at rest, decrypted it in memory, and used server-side ownership checks to revoke access after a sale.
the technical implementation worked, but market behavior showed that buyers valued crypto-native trading more than the ownership and display experience we designed. artists valued file protection less than expected, and the two-sided marketplace was harder than either of us anticipated.
the biggest lesson was that we had solved sophisticated technical problems before validating whether they were the market’s highest priorities. that experience substantially changed how quickly i try to expose product assumptions to reality.