Neuronix

Overview

Neuronix is a WordPress port of a client-supplied Gemini/AI Studio React prototype for an IT technology blog and knowledge hub — but unlike this catalogue’s usual faithful-port builds, the brief for this one explicitly called for auditing and fixing the prototype’s functional logic, not just carrying its visuals across. Before any code was written, the full ~6,500-line source was read end to end, surfacing eleven real functional defects hiding behind a polished UI: a KV-cache memory calculator missing the layer-count dimension, a sound-effects toggle whose default “on” state was never actually wired to the audio engine, a “force-directed” knowledge graph with velocity fields that were never integrated, an AI reading assistant that returned the same two canned answers regardless of the question, reading-preference settings that silently reset on every reload, and more.

1. A real, content-grounded AI reading assistant, not keyword-matched canned answers. The original component matched two fixed keyword sets and returned one of two hardcoded paragraphs no matter which article or question a reader typed. The rebuilt version scores query tokens against that specific article’s own structured takeaways and table of contents, and gives an honest “not found in this article” fallback when nothing matches well enough — it never pretends to be a general-purpose model.

2. Real force-directed graph physics for the Knowledge Graph page: pairwise node repulsion, spring attraction along edges, and a centering force, integrated frame-by-frame — replacing the prototype’s hardcoded pixel positions that only looked physics-driven.

3. A corrected KV-cache memory calculator on the Interactive Code Lab page: the original formula silently dropped the number of transformer layers from the computation and treated a “K tokens” context-length label as a literal token count without the x1024 conversion, producing numbers that looked plausible but weren’t real. The fixed version was hand-verified against MHA/GQA/MLA reference figures for DeepSeek-V3’s actual architecture.

4. Real, durable WordPress comments and newsletter capture in place of the prototype’s component-state-only versions, which silently discarded every visitor comment and every submitted email on refresh — both now persist for real (comments through core’s own `wp_handle_comment_submission()` pipeline; newsletter signups to a dedicated DB table with honest, non-overclaiming success copy since no outbound mail service is wired up).

Independently re-verified after the build agent’s own report: caught and cleaned up WordPress’s own default seed content (a published “Sample Page” and a draft “Privacy Policy”) both still carrying the literal dev-host address baked into WP core’s boilerplate text — the same address-contamination bug class independently caught in this catalogue’s SinoChemix build. Also found, root-caused, and fixed a silent total failure of the AI assistant for Chinese-language questions: the tokenizer only split on ASCII whitespace, so any natural Chinese sentence (the primary language of this entire site) collapsed into one unmatchable token, while a bare English term happened to work — fixed with a CJK bigram-splitting tokenizer, then re-verified that a previously-failing on-topic Chinese question now answers correctly and a deliberately off-topic Chinese question still returns the honest “not found” fallback rather than over-matching.

Bilingual, English-first (v1.2.0). The whole front end runs in two languages from one install: English is the native source text and the Chinese lives beside it — gettext for theme copy (386 catalogued strings, .pot/.po/.mo shipped), and per-record meta for everything gettext cannot reach: post titles and bodies, structured article meta, taxonomy names and descriptions, menu labels, author names, job titles and bios. Every runtime string in the JavaScript is routed through wp_localize_script, so the reader, code lab, tech radar and knowledge graph all speak the visitor’s language too. A header switch sets a year-long cookie and redirects to a clean URL; responses carry Vary: Cookie, hreflang alternates and the right lang attribute (zh-Hans, not zh-CN), and Chinese pages load a CJK font stack. Category, IT-directory, tech-radar and knowledge-graph content is editable from wp-admin in both languages.