I find the right ways to apply technology to real estate.

I won 2nd place with my good friend Valerie Yong at the Rosewood Sand Hill Hospitality 2030 Hackathon, organized by Cerebral Valley. The event, sponsored by Rosewood, Anthropic, ElevenLabs, and Greycroft, focused on how luxury hotels can apply AI to improve and refine guest experience across a global portfolio.

Our core product principles were:

1. Don't tax the customer with forced data intake (paperwork) or web scraping; always solve for creating an effortless guest experience.

2. Luxury hospitality is a global business supported by international teams. In a highly personal, people-first setting, AI can remove language barriers to improve staff choreography, enabling each team member to deliver the highest-quality experience with every interaction.

A sweet full-circle moment: I entered and won my first hackathon a little less than a year ago, and have placed in seven total since — each an accelerated step on the path to becoming technical, working alongside experienced engineers to ship my own ideas fast.

Ask a room of institutional investors where San Francisco rents are headed and the answers range from 2.5% to 12%. HomeStar doesn't forecast rent directly. It watches hiring instead — job postings and labor flows, which move months before rents do.

I tested it against the obvious alternative, forecasting rent from past rent, and it held up better on cities it hadn't been trained on. Every input is public: job postings and published labor statistics, collected through Bright Data. First place in the Finance track with Salim Masmoudi.

Most agent-commerce demos only run one way: agents paying agents. Gyasss runs both ways. People earn USDC for confirming gas station prices, and AI agents pay a tenth of a cent per query to read them.

The hard part is trust — paying for crowdsourced data invites people to game it, so contributions are weighted by agreement, decay over time, and can be staked and slashed. Sub-cent payments only work on rails built for them: the same transaction costs a thousand times more in Ethereum gas, and Stripe's 30-cent minimum is 300 times the price. Built solo over a week.

MARA Holdings is the second-largest corporate holder of Bitcoin. They challenged participants to build an AI-driven system to arbitrage energy and AI inference prices across their global data center portfolio.

My team built Panoptic, a platform that integrates satellite data with real-time energy pricing to optimize compute allocation and infrastructure investment. The system includes a global visualization of MARA's data centers, AI-powered analysis for battery storage buildouts and expansion siting, an automated energy treasury that deploys power based on grid pricing, and a derivatives platform for hedging Bitcoin and energy exposure.

The derivatives component drew directly on my background in interest rate risk — the same hedging frameworks that work for real estate debt also work for energy and crypto treasury.

Traditional A/B testing is slow and manual. I designed a system that applies evolutionary biology to user interfaces — beneficial changes propagate, harmful ones die off, and the UX evolves autonomously.

Survival of the Feature uses two AI agents working in coordination. The first generates UX variations and deploys them to a sample of users. The second applies statistical analysis to real-time user response, determining whether each variation improves engagement while calibrating how quickly to expand exposure. It's similar to natural selection, except data accelerates the process.

We won first place out of 45 teams and presented at the conference keynote.

My first hackathon. Roof inspections are expensive, slow, and miss anything you can't see from a ladder. Roofi pulls frames out of drone footage and runs them through a vision model to flag cracks, missing tiles, leaks, and debris — hundreds of images from a single flight.

The pipeline is Python: ffmpeg extracts frames at configurable intervals, batches them through Google Gemini's vision API, and consolidates the responses into JSON that downstream reporting can pick up. A person on a ladder sees one elevation at a time; this sees the whole roof.

First place out of more than 50 teams at the Aparavi hackathon.

A construction consultant's site visit is the most informed hour a project gets — but it only happens periodically. Between visits, the owner and lender are working from the last report, and a problem that surfaces in week two waits until week four to be seen.

SpotCheck covers the gap between visits. A multimodal AI agent analyzes daily drone footage, builds 3D models of the site, scans for OSHA compliance issues, and flags schedule slips early. You can ask it questions in plain language — how a specific area is progressing, what changed since last week. The consultant arrives already knowing where to look, and the owner and lender see the same current picture the consultant does.

Built at the AWS MCP Agents Hackathon, which required chaining at least three sponsor tools into one agentic workflow: AWS, Minimax for video understanding, BEM for 3D reconstruction, Operant for agent security. Third place out of more than 50 teams, one week after Roofi — our second drone project in two weeks.