A personal site for side projects and ad hoc explorations.
Projects
Lectorate
Readability tools tell you a grade level; I wanted the question that actually matters for public writing: how many people does this exclude? Lectorate is a machine-learning model I trained locally on 111,000 pairwise human difficulty judgments (the CLEAR corpus), then chained to real population data — the PIAAC 2023 adult literacy survey, UNESCO literacy rates, Ethnologue speaker counts — so any pasted text returns the share of US adults and the world's adults who could functionally comprehend it, with honest conformal error bars. The version here is the exact trained model exported to JSON and re-implemented in dependency-free JavaScript, parity-tested against the Python original, running entirely in your browser. Try the same benefits letter written two ways: the gap is about 100 million American readers.
San Francisco — A Private Guide
A personal city interface I built for my mother-in-law's month in San Francisco around the birth of my son: simple enough to share as a URL, opinionated enough to feel like my guide. Instead of spending hours wrangling Google Maps lists, I used voice prompting with a coding agent to turn favorites like Rose's polenta, Turtle Tower, Dalida, the Legion of Honor, and the de Young into a lightweight map with search, filters, and one-tap navigation. It became a small proof point for a bigger belief: AI will let more people turn their taste into personal software.
Shop Lens
An imaginative-shopping prototype for turning an existing scene into a counterfactual one. Upload a room photo, describe the party or room you want to create in natural language, and the system uses Gemini 2.5 Pro to infer intent, query SerpApi for real inventory, and pass the product assortment into Gemini 2.5 Flash Image to recompose the scene with items you can actually buy. I built it because I love throwing themed parties but hate the search-and-revise loop of online shopping; the hard product work was giving the agent low-latency inventory and iterating until the generated images and shoppable items matched the user's taste.
Night Sky
A private, browser-based sky map inspired by childhood nights at Hetch Hetchy, where a naturalist once led us out at 2 a.m. to lie under the Milky Way during a meteor shower. I ran the same prompt across Claude Code, Codex, and Antigravity to evaluate one-shot constellation accuracy, time-to-output, and qualitative UX, then productionized the strongest result into a tool that accepts real addresses and computes constellations, planets, the sun, moon, and moon phase in the browser. The learning was both technical and personal: AI can produce surprising scientific interfaces, but the craft is in making complexity legible and trustworthy.
Baby Name Rater
Disposable software I built while my wife and I were choosing our son's name: a small, purpose-built app for a temporary but meaningful decision. The problem was not a lack of baby-name lists; it was that tone, timing, cultural constraints, and busy schedules made preference-sharing noisy. The Hinge-like rating and matching UX gave each of us a private login, a fast flow through names, meanings, and origins, partner-aware ranking, and a shared shortlist when we both liked something. It taught me that software can structure an emotional decision without pretending to replace taste — fittingly, the name we chose, Avi, was on the list.
Contact
Ask about Noah's experience or send him a message.
Hi there! I'm Noah's assistant. I can tell you about his professional background, or help you get a message to him. What can I help you with today?
Responses are AI-assisted. Do not submit sensitive information. Conversations may be logged and retained up to 5 years. By sending a message, you consent to Noah receiving your contact information.