Founder,
building agents.

Let me back up.

1,000+ people applied. 30 builders were selected. The Voice & Video AI Hackathon at HF0's mansion in San Francisco — hosted by Knowtex, NomadicML, and HF0 — only wanted exceptional builders. I'm a sophomore at UMass Amherst. I almost didn't go because I couldn't afford the trip. I went anyway.

12 hours. Unlimited credits from Anthropic and OpenAI. A room full of elite engineers building with swarms of AI agents, multimodal pipelines, and bleeding-edge voice and video models. The energy in that mansion was unreal.

We built ColdBrew: a platform that gives warehouse cameras a brain. Real-time detection of safety violations, equipment failures, and shipment anomalies using NomadicML's vision API. We demoed it in front of top VCs, founders, and AI engineers in one of the most exclusive rooms in the Bay Area.

First place. Won a Mac Mini. Also may have pitched a few VCs and founders about Assemblr (building with Aditya Hemanth Vellanki, launching soon) between rounds. You don't fly cross-country and not work every angle.

If you're a student reading this and you think you can't afford to be in these rooms — find a way. Book the flight. Figure it out on the way there. The worst thing you can do is not show up.

But the prize isn't the part that mattered. The part that mattered was being 20 years old in a room at the Archbishop's Mansion with engineers from Roblox and xAI, founders building at HF0, and some of the sharpest people I've ever met. Being one of the youngest in the room and feeling like I belonged.

Huge shoutout to my teammates Nisarga Patel, Mohammed Adnan, and James (Jaime) Orellana Orellana — we met that morning and shipped a winning product by evening. And special thanks to Mustafa Bal, Caroline Zhang, Jocelyn Kang, Evan Stites-Clayton, Akshay Gopalkrishnan, and Aadit Bhatia for putting together an incredible event. Great connecting with Auriel Wright, Nate Stone, Andor Kesselman, Shwetha Rao, and Rudresh Upadhyaya among many others.

I fly back today. Still processing. One of the best experiences of my life. Can't wait to come back to SF.

Most of my work is in agentic AI and developer tooling, so this was a fun exercise in cross-domain applicability. Turns out the techniques carry over pretty well.

We benchmarked GPT-4.5, Gemini-2.5, Claude-3.7, and DeepSeek-R1 against the WCRF/AICR evidence-based dietary guidelines, comparing model outputs against human responses across adolescents, adults, and elders.

Excited to share that the work was just accepted for presentation at NUTRITION 2026 — ASN's flagship annual meeting.