03 · Demos
AI
Demos
Move past the chatbot. Hands-on explorations of what LLMs can actually do when you understand the architecture.
RAGIT — NYC Mayoral Race (October 2025)
Mamdani vs. Cuomo: candidate positions visualized in semantic space
With the NYC mayoral elections on the horizon and a lot of hype to sift through, I built an app for quick, intuitive candidate comparison. RAG on each policy document → semantically meaningful vectors → graphed and clustered → voters can compare positions on Housing, Energy, Healthcare, and more at a glance.
My favorite feature: the tooltips. Zoom in on any vector dot and instantly see the text behind it. Frontend built on Replit. Vibe coding is fun — but weak on clustering and guardrails, as it turns out.
See the Full Post on LinkedIn →
SEC Filing Dashboard
Document intelligence over financial filings — built fast, shipped honest
A Replit-built dashboard for exploring SEC filings. The core demo came together quickly — then Replit Agent offered to exceed expectations by weaving in sentiment analysis. It stumbled on NLTK punkt. I tried pointing out the bugs; Replit couldn't do a quick recovery.
That enhancement wasn't on the critical path for the mockup, so we time-boxed it to 10 minutes and moved on. An honest vibe-coding story: fast on the core, fragile on the edges.
Campus Whimsy
A fun and transformative app for college discovery and admissions guidance
A proof of concept exploring what AI-powered college counseling could look like for high-achieving students from academically elite public schools — the ones who don't have access to $300/hour advisors. Designed as a mobile-first experience for 9th and 10th graders.
Built on Replit. This demo is about what the initial user journey could feel like — college location, majors, admission requirements, school stats — surfaced through a conversational, relatable interface.
Campus Whimsy is an exploration, not a product. What it demonstrates is the full stack of skills behind it: AI product strategy, rapid prototyping, user-centered design, and a clear thesis about where GenAI can create real equity in education.
Backpropagation, Live
Teaching a simple network to learn the XOR function — and watching it work
XOR is the classic test case for non-linear learning: no single neuron can solve it, but a tiny two-layer network can. Watch the decision surface reshape in real time as backpropagation adjusts the weights, epoch by epoch. Theory coming to life in just a few lines of code.
Experiment with the learning rate and see how it affects convergence. Want to try it yourself? Grab the HTML artifact from GitHub and run it locally — or feed it to Claude to swap in a different activation function.