niltoidA dev blog and project portfolio
AboutAgent view

Ahmad Saeed · Product Engineer

About

Ahmad Saeed builds things that respond: robots and LED rigs, interactive installations, games for activism, generated visuals, and AI tools. This site is a scrapbook of the projects he’s explored over the years.

The etymology of “niltoid”

A made up word that could mean “resembling absolutely nothing” or “having the characteristics of zero.”

Nil-
Latin nihil: “nothing” or “zero.”
-toid
From Greek-derived -oid: “resembling” or “having the form of.”

Music projects

How this site works

The Ask bar is meant to serve as the main navigation model for this site. Type in a natural language query and receive the posts that are relevant matches.

To accomplish this, this site leverages Jev, a System 1 model built for fast decisions, in contrast with the more deliberate System 2-style processing of reasoning LLMs. It accepts text context and returns one constrained answer per question. Three return types are supported: Noul (a yes/no probability), Choice (from given options), and Score. This limitation allows it to perform incredibly quickly and cheaply for the right kinds of problems. When a query is submitted, we run it against every article in one batch and receive a Noul response for each one. Local rules choose the layout from the query: editorial for general topics, a timeline for chronological questions, or a table for comparisons and skills.

A built-in keyword matcher also searches titles, summaries, tags and skills. If Jev is unavailable or takes too long, or if I spend all my Jev credits, the site uses those results to keep the Ask bar working.

Category and layout suggestions have precomputed answers: prepared selections of posts with a fitting layout, shown immediately. Jev and the keyword matcher also evaluate those prompts in the background so I can compare their results and improve the search. Suggestions that name a specific article open it directly.

From question to page
Ask barA typed question or suggested prompt
Article catalogWritten posts, code and metadata, prepared at build time
JevThe Cloudflare Worker asks Jev to score each post against the question.
Keyword matcherThe browser also matches titles, summaries, tags and skills.
Precomputed suggestionsCategory and layout prompts show prepared answers immediately; named-article suggestions open the article directly.
Choose the answerShow prepared editions while comparisons run in the background. For other questions, use Jev, with keyword results as the fallback if Jev is unavailable.
Editorial · Timeline · TableThe page arranges the matched posts. If nothing matches, it offers front-page picks.
PostHogAlongside the page: questions, both rankings, displayed posts and article clicks.