How this site works

A portfolio is read far more than it is used. The plan asked for a Lighthouse score of 95 or more on every page, a largest contentful paint under 1.5 s and small JavaScript budgets.

Astro builds every page to static HTML. Preact runs only in islands that need it: the assistant, the fit check, the status page and the search palette, which loads on first use.

Every page is complete without JavaScript. CI fails a page whose initial JavaScript goes over its budget: 60 KB gzipped on the home page, 40 KB elsewhere. Islands render at build time first, so hydration moves nothing on screen.

Pages and the API share one origin, so the browser needs no CORS. A Worker invocation costs CPU time, and almost every request is for a page or a file.

Cloudflare's asset layer serves the build output directly. The Worker runs first only for /api/* (run_worker_first), with Hono routing inside it.

Pages never wake the Worker, and the API never sends CORS headers. Page headers live in a _headers file. The Worker sets its own on API responses, including a Content-Security-Policy that lets nothing load.

The assistant answers only from this site's own content, a small set of Markdown entries. A hosted vector database or file search would add a service, a network hop and a bill.

A build script splits the public entries into passages, embeds them at 512 dimensions, stores the vectors as int8 and commits the index as JSON. The Worker bundles it and searches it in memory.

Retrieval needs no network call besides the question's embedding, and every content change shows up in review as an index diff. A unit test fails when the index is stale. It would not scale to millions of passages, and it does not need to.

Names and acronyms such as DAL-A need exact matching. Paraphrases, and questions asked in Spanish or German, need meaning. Neither signal is enough alone.

Rank the passages twice, with BM25 and with cosine similarity, drop weak matches from each ranking, and merge both with reciprocal rank fusion (k = 60). The top 5 go into the prompt.

Recall@5 reached 95.7% on the golden set. Fusion only sees ranks, so the relevance floors apply before it. When the embedding call fails, BM25 still answers, and suggested questions carry precomputed embeddings.

The assistant is public and every answer costs money. There is no login and no cookie, and any counter kept in the browser can be edited.

Cloudflare Turnstile checks the browser once, then the Worker issues a signed, anonymous visitor token. After a burst limit, the server counts each question in one D1 transaction against the visitor's 10 and against daily caps per network and for the whole site.

Concurrent requests cannot overshoot a limit, and editing the counter in the browser gains nothing. Clearing storage starts a new allowance, but the network and site caps still hold. A question is given back only when the model fails before its first word.

A visitor judges an answer by how soon it starts and by whether it can be checked.

Answers stream as Server-Sent Events from a plain POST. The first event carries the sources, the text follows in deltas, and a done or error event ends the stream.

Source links appear before the first word. The client treats a stream without done as failed and offers a retry. When the visitor leaves, the Worker aborts the model call, so no more tokens are billed.

No prompt can promise that a model never repeats its instructions. In the red team, a paraphrase and an encoded reply both got past the prompt alone.

The instructions carry a secret canary, and every delta passes a streaming guard before it reaches the visitor. The guard holds back a short window and stops the answer on the canary, on the instructions, verbatim or retold, or on a long encoded run. It also strips HTML and links to other sites.

An answer can stop mid-sentence, and that question still counts, so tripping the guard never buys a free question. The guard runs on every delta, so it scans only the new text to keep CPU time low.

A job post is untrusted text and may carry instructions for the model. What the model returns is untrusted too.

Two model calls with strict JSON schemas extract the requirements, then propose a level and evidence for each. Deterministic code then keeps only evidence retrieval offered for that requirement, requires corroborated retrieval for demonstrated, applies a years rule, and caps every level at related when the post addresses automated tools.

A fully steered model can at worst produce related, backed by real, retrieved evidence. The price is some recall: an honest match the rules cannot corroborate stays at related.

The Workers Free plan allows 10 ms of CPU per request. Parsing about a thousand streamed events per analysis cost most of the fit check's CPU, and Cloudflare cut analyses off.

Both fit check model calls ask for the whole answer. The visitor still gets the rows over Server-Sent Events, each one sent after validation. The chat keeps streaming.

The fit check's CPU time fell from 341 ms to 15 ms at p50. The first row now waits for the whole classification instead of arriving as soon as the model writes it.

A portfolio has no reason to load trackers, and every third-party script widens the attack surface. Some visitors will open the developer tools.

Every page carries a Content-Security-Policy with a SHA-256 hash for each inline script, and no unsafe-inline. Stylesheets and fonts are files on this domain. Turnstile, the only third party, loads on the first focus of the question box or the job post field.

A build check fails on an unhashed inline script or a style attribute. Mozilla Observatory grades the site A+. Adding analytics later means changing the CSP and the privacy notice together.

In October an audit of the assistant found it joining true facts into false ones: a client of the real estate platform became a client of his employer. Retrieving a few passages per question let the model fill the gaps, and the same question got different answers.

The model reads all of Tomás's published statements and answers as structured output, each sentence naming the statements it relies on. Code checks every figure, date, name and relation against them, and a second model removes any sentence they do not support. Common questions get a prepared answer, served word for word. The cards under an answer come from the site's own data.

An answer can only say what Tomás has published, and the visitor can open the statement behind each sentence. A sentence that fails a check is retried once, then removed or replaced by an honest "not listed here" with his contact details. The second check adds about a second. It runs on the preview first, and production switches once the audit passes.