The code, and where the rest lives.
Every shipping claim on this site sits next to the thing shipped. This page is the index: what is public, what is private by employer, and the engine that answers questions on the homepage.
This site, including Brix
Plain HTML, CSS, and JavaScript with no build step. The retrieval engine, the knowledge base, the accessibility architecture, and the QA tooling are all in one public repository.
Global Content Studio Financial-Finesse/ff-content-writer
TypeScript and Lit on Vercel, a Cloudflare Worker that proxies the Anthropic API and fronts Azure Blob Storage, and the Atlas localization pipeline. 524 commits, 493 of them mine. 221 Vitest tests across 12 files.
Hub Content Studio and localizer review queue Financial-Finesse/blue
The content authoring and localization review surface inside the company's core Hub application. TypeScript and Lit, 80 commits, with tests for the markdown, metadata, adjudication, rail, and library logic.
Localization review API and console Financial-Finesse/ff.workers
The Python review API and core, its tests, and the React review console the localizers work in. 27 commits.
Aimee SDK Financial-Finesse/benehub
Engineering built the SDK in React and A2UI from my coded prototypes and spec. I then merged the design-token layer and the empty-state, conversations, icon, and card-shell components into it. 32 commits, July to August 2026.
retrieve(), verbatim
This is the function that answers questions on the homepage, copied from index.html. It is a keyword engine, not a model: normalize, drop stopwords, expand synonyms, fuzzy-match against a hand-written knowledge base, and score. If this and the homepage ever differ, the homepage is the truth.
// index.html — the engine answering you right now (verbatim)
function retrieve(query) {
const nq = norm(query);
const qt = toks(query);
let best = null, bestScore = 0, second = null, secondScore = 0;
const scored = [];
for (const e of KB) {
// `not` keys veto an entry outright — this is what stops "case study"
// scoring against the education entry via the word "study".
if (e.not && e.not.some(n => nq.includes(n))) continue;
let sc = 0;
for (const key of e.keys) {
const kl = key.toLowerCase().trim();
// A multi-word key is a much stronger signal than a single word,
// so it outranks a generic one-word match on another entry.
if (kl.length >= 3 && nq.includes(kl)) { sc += 3 + (kl.split(' ').length - 1) * 2; continue; }
for (const w of kl.split(' ')) {
if (!w || STOP.has(w)) continue;
if (qt.some(t => fuzzy(t, w))) { sc += 1; break; }
}
}
if (sc > 0) scored.push({ e, sc });
if (sc > bestScore) { second = best; secondScore = bestScore; best = e; bestScore = sc; }
else if (sc > secondScore) { second = e; secondScore = sc; }
}
scored.sort((a, b) => b.sc - a.sc);
return { entry: best, score: bestScore, second: secondScore >= 2 ? second : null, scored };
}