CV Viva's chat answers in Spanish or English about Henry Rincón's career, always in the third person and only with the site's content: the résumé, 25 in-depth documents, every Portfolio piece and the Methods publications. Every claim carries its citation to the section that backs it, and if the AI provider goes down, it keeps answering with local search.
An answer with its citations: the chips lead to Vesting and BI & decision«What are his weaknesses?»: the fixed text, with no sourcesWith no provider, local search shows the fragments that answerThe chat's gate: name, email and the Law 1581 notice
Chat features: One per chat feature documented in its user manual, matched section by section: 18, in six groups.Index fragments: The citable pieces the site's content is split into so the chat can search them: 1939 in Spanish and 1930 in English, counted when the site was built on October 31, 2026.Bank questions, per language: Questions written in a recruiter's words, the same 175 in each language. Per language, the bank also holds 15 unrelated ones, 1 that gets refused and 3 declared gaps.Switchable providers: The AI services chosen with a variable: Groq, Gemini, Azure with Microsoft Foundry, Anthropic and an OpenAI-compatible one. The tests' simulated one doesn't count.Tokens in fixed answers: By construction: the refusal, unrelated questions and leaving Vesting are answered on the server before a provider is chosen, so they never call the model.
in-depth documents within reach: the site doesn't publish them, the chat reads and cites themmeasured
Besides the résumé, the Portfolio and Methods, the chat reads 25 in-depth documents Henry wrote and the site doesn't publish as pages: all 25 are approved and indexed in Spanish and English. A citation doesn't open the document: it says where the sentence came from (AF-09, for example) and leads to the section of the site that backs it.
For recruiters who want a specific answer (“what did he do at Vesting?”, “what does he know about Microsoft Fabric?”) without reading the whole site, and for anyone who wants to check it: every answer says where it comes from and takes you there. What isn't on record isn't answered, and the chat says so.
A résumé chat that doesn't make things up takes more than a good prompt. Here, what isn't on record (what he didn't do, what he doesn't know, salary, a credential he doesn't hold) gets a fixed text before it reaches the model, without spending a token; leaving Vesting has its own fixed answer with its citation, and the panel shows only the sources the answer cites. Behind it sit an index rebuilt on every publish and a bank of 175 outside questions per language that proves, on every change, that the content answers. And it never goes silent: if the provider fails, the same search runs in the browser.
Top to bottom, in the order a question happens: the visitor comes through the gate, asks and gets an answer with its source or a fixed text, and the chat stays up even if the provider fails or the page reloads. On the right, what runs through everything.
One chapter per group: what it solves, when it fits, how it works, its screens and its cards.
You can't ask a résumé anything, and an answer without a source can't be checked. Here whoever asks gets two or three paragraphs in the third person, built only from the site's content, and every claim carries the number of its source, which leads to the section that backs it.
Whoever asks gets two or three paragraphs based only on the site's content, and every claim ends with its citation. The chip says which source the sentence came from and which section it leads to, without losing the conversation.
An answer with its citations: the chips lead to Vesting and BI & decisionThe chat talks about Henry, never as Henry. The in-depth content is in the first person because it is his voice, and the answer moves it to the third; a test flags the owner's voice if it slips in.
It runs through the whole app.The index adds every Portfolio piece and the Methods publications, block by block: a question about an app, an agent or a method brings its citation, which leads to the app's brochure, the piece's sheet or the exact block.
The answer cites the Profile, Vesting and a Methods publicationWhat does a chip like “[1, 2] AF-09 · Vesting” say?
The numbers are the marks in the answer that come from that source. The code says where the sentence came from: AF for an in-depth document, CV for the résumé, APP for the apps list, FT for a Portfolio piece, and the publication's number for Methods (T4, for example). The last part says where it leads: the section of the site that backs the sentence.
Who is speaking in the chat?
The chat may talk about itself, but it always talks about Henry in the third person. His in-depth documents are in the first person because they are his voice, and the answer moves them to the third.
It may say “I only answer from what the résumé actually shows”: it talks about itself, not about Henry.
Always “Henry”, “he”, “his”; never “I” or “my” for his experience. A test flags the owner's voice if it slips in.
You ask about a Portfolio app: what wins, what the Portfolio publishes or Henry's documents?
If the question is about a piece, what the Portfolio publishes wins and the citation leads to its page; if it is about Henry, his document wins. In the search, Portfolio fragments weigh half as much as in-depth documents: they are evidence of one piece, and Henry's voice is his documents.
The answer uses two parts of the same document: do two chips show up?
No. One chip shows up with both numbers, like “[3, 4] AF-17 · BI & decisión”, and on hover it shows the titles of both parts. No number is lost, because the answer cites by number.
A model asked about what isn't written down may fill the gap. Here absences, salary, availability or an unrelated topic get a fixed text before reaching the model, without spending a token, and leaving Vesting has its own fixed answer, with its citation.
“What has he never done?”, his weaknesses, a credential he doesn't hold, salary or availability all get the same fixed text, in the page's language, with no sources and without calling the model.
«What are his weaknesses?»: the fixed text, with no sources“Why did he leave Vesting?” has a fixed answer that tells how the stage closed, with its citation to the Career section. It adds no causes the content doesn't give and spends no tokens.
«Why did he leave Vesting?»: the fixed answer, with its citationThe weather, a joke, a recipe or a request for code get a friendly fixed answer that invites a question about the career, with no sources and without calling the model.
A joke gets no answer: the chat invites a question about the careerWhat does “not on record” mean?
What the résumé doesn't say: what Henry didn't do, doesn't know or doesn't have, a credential that isn't listed, salary, availability or whether he travels. The content isn't the absolute truth and something may slip through, so the chat never claims an absence: it gives one fixed text, in the page's language, and invites you to write from Contact.
“What doesn't HackGuard do?”: does that get refused too?
No. A piece's limits are written in the Portfolio and get answered; what gets refused is the person's absences. And “what does he do when a role asks for something he hasn't used?” gets answered too: it is his method.
“What has he never done?”, his weaknesses, a credential he doesn't hold: fixed text, no sources.
“What doesn't HackGuard do?”: its limits are on record in its brochure, and the answer cites them.
You ask “Why did he leave Vesting?”: what does it answer?
A fixed answer, in the third person, that tells how the stage closed in January 2025, with its citation to the Career section. It adds no causes the content doesn't give, doesn't call the model and works in local search too. Two prompt rules in a row didn't get the model to answer only with what's on record; that's why this question no longer reaches it.
How does the chat know a question is unrelated?
In two ways: a closed list (the weather, a joke, a recipe, a request for code) and a strict search of the index that finds nothing. Either way it gives a fixed invitation to ask about the career, with no sources and without calling the model. Whatever slips through reaches the model, and the prompt declines it.
A chat that depends on a single AI service goes quiet when that service fails or runs out of quota. Here it switches to local search over the same index and says so; the provider is chosen with an environment variable, without touching code, and another variable turns the whole chat off.
If the provider goes down or runs out of quota, the chat switches to “Local search” and shows the fragments that best answer. After three failures in a row it stops trying for a minute.
With no provider, local search shows the fragments that answerGroq, Gemini, Azure with Microsoft Foundry (formerly Azure AI Foundry), Anthropic or a self-hosted OpenAI-compatible server: switching provider is switching an environment variable, without touching code.
It runs through the whole app.One variable turns the whole chat off: the button disappears from every page and the server route answers that it is off.
It runs through the whole app.What is “Local search”?
It is the fallback mode. When the AI provider fails, the browser downloads the same index the server uses and shows the four fragments that best answer, trimmed and numbered, with the “Local search” label in plain sight. It doesn't write: it shows the evidence as it is. The index is downloaded only at that moment.
Model answer or local search?
Both come from the same index and the same number of fragments; what changes is who puts the answer together.
Two or three paragraphs in the third person, with their [n] citations, built from the four fragments.
The same four fragments, unwritten, numbered and with their chips; the fixed texts still answer too.
The provider retires the model the chat uses: what happens?
Answers keep coming, in local search and with the label in plain sight. The automated tests can't see it, because they use a simulated provider; the owner notices it in the label or in the server logs, and fixes it by choosing another model with a variable, without touching code. It has happened once, and the chat didn't go silent.
The owner wants to turn the chat off right now: how?
He sets the switch variable to false and republishes: the button disappears from every page and the server route answers that the chat is off. To turn it back on, he deletes the variable or sets it to true.
An open chat can be used up by anyone, and on the other side nobody knows who asked. Here the visitor leaves their name and email, accepts the Law 1581 notice and gets in with a code; each person has 20 questions every 24 hours, an email can be blocked and every conversation goes into a log only the owner reads.
Before the first question, the visitor leaves their name and email, accepts the Law 1581 notice and receives a six-digit code valid for ten minutes. With it they get in and don't register again for 30 days.
The chat's gate: name, email and the Law 1581 noticeEvery question is logged with its answer, its sources and its mode (AI, fixed text or local search), and only the owner reads it. Anyone who wants it deleted asks from Contact.
It runs through the whole app.Each email gets 20 questions every 24 hours, counted in the log, and whoever hits the cap is asked to come back tomorrow. An email on the blocklist gets no code and can't ask.
The daily cap: you reached today's maximum of questionsIf something fails, “Something not working? Tell Henry” sends a report with an optional note. Inside the chat it doesn't ask for the email: the server takes it from the session.
«Tell Henry», opened inside the chat without asking for the emailWhy does the chat ask for a name and email?
So Henry knows who is asking and can follow up, and so someone who only comes to poke around doesn't use up everyone's chat. There are no passwords or accounts: the six-digit code proves the email belongs to whoever typed it, and with it the session lasts 30 days in that browser.
The daily cap or the per-minute rate?
They are two different brakes: the quota counts the person; the rate, the connection.
20 questions every 24 hours per email, counted in the log: it holds across every server and is exact.
10 questions per minute per connection, in each server's memory: it stops bursts, but it isn't exact.
You've already asked 20 questions today: what happens to the next one?
You see that you've reached today's question limit: come back tomorrow or write to Henry from Contact. It isn't a failure, so the chat doesn't switch to local search. Every question counts, including the ones that get a fixed text.
Something isn't working inside the chat: how do you report it?
With “Something not working? Tell Henry”, at the foot of the panel: you write what happened, if you want, and the report reaches him by email. Inside the chat it doesn't ask for your email: the server takes it from your session and never sends it back to the browser. If what fails is the gate itself, the same report is on the first step.
A conversation that's lost on reload or a language switch forces you to start over. Here it stays in the browser when you reload, follow a link or switch language; it expires by itself 8 hours after the last message and “Delete conversation” removes it with one click. All of it, in Spanish and English.
The conversation survives a page reload, following a link or switching language. It clears itself 8 hours after the last message, and “Delete conversation” removes it from that browser.
Conversation cleared from the browser; the server log is keptThe panel, the fixed answers and the sources come out in the page's language, and the refusal lists read both languages, whatever the page's language.
It runs through the whole app.Browser or server?
They are two copies with different purposes: the browser's is for you; the server's, so Henry knows who asked and what was answered.
Up to 40 messages with their sources, without name or email. It expires 8 hours after the last message and you clear it with one click.
Every question with its answer, sources, mode, your name and your email. Only the owner reads it, and it is deleted on request from Contact.
You open the site in a private window: does the chat work?
Yes, the same, but without memory: if the browser doesn't allow storage, the conversation doesn't survive a page reload.
You write in English on a Spanish page: what happens?
The chat answers in the page's language. The refusal and leaving-Vesting lists read both languages, so a question about salary written in English still gets the fixed text.
A chat that searches by words only finds what the content says in the asker's words. Here the index is rebuilt on every publish, only from what's approved, and a bank of 175 questions per language, written in a recruiter's words, requires on every test run that each one brings the source that should answer it.
Every publish rebuilds the index from the résumé, the in-depth documents, the Portfolio and Methods. A draft document doesn't go in, and an approved one requires its English twin.
It runs through the whole app.Every test run asks 175 questions per language, in a recruiter's words, and requires each one to bring the source that should answer it.
It runs through the whole app.Continuous integration uses a simulated provider. A separate pass, only on the owner's machine and with his permission, asks the guide's questions to the real model and marks every answer.
It runs through the whole app.Question bank or live pass?
The bank proves the content answers; the live pass, what the real model does with it.
175 questions per language on every test run, with a simulated provider: it checks that the search brings the source.
The guide's questions to the real model, only on the owner's machine and with his permission: it marks voice, absences, motives and sources.
A bank question turns red: what's done?
There are two honest ways out, and neither is loosening the test: if the content doesn't say it in the asker's words, the content is fixed; if another source answers it better, the expected source is fixed. Making up content so it passes is forbidden.
What goes into the chat's index?
The résumé, the apps list, the approved in-depth documents, every Portfolio piece and the Methods publications, block by block: 1939 fragments in Spanish and 1930 in English at the last count. A draft document doesn't go in, and an approved one requires its English twin.
What happens to the questions the content doesn't answer yet?
They are declared as gaps and never made up: today there are 3 per language (salary and availability to start or to travel), and each waits for something only Henry decides. Meanwhile they get the fixed text. The test counts them and caps them: the list can only shrink.
The whole path. No swimlanes: every box says who does it, and the ones with the screen icon show it.
Every figure says where it comes from. And what the app does not do is spelled out too.
One per chat feature documented in its user manual, matched section by section: 18, in six groups.
The citable pieces the site's content is split into so the chat can search them: 1939 in Spanish and 1930 in English, counted when the site was built on October 31, 2026.
The documents Henry wrote for the chat, beyond the résumé. All 25 are approved and indexed in both languages.
Questions written in a recruiter's words, the same 175 in each language. Per language, the bank also holds 15 unrelated ones, 1 that gets refused and 3 declared gaps.
End-to-end chat journeys in a real browser, 47 on desktop and 47 on mobile, with a simulated provider.
The AI services chosen with a variable: Groq, Gemini, Azure with Microsoft Foundry, Anthropic and an OpenAI-compatible one. The tests' simulated one doesn't count.
How many fragments the model receives, numbered, to build each answer and cite them. It's a number set in the code, and local search shows the same four.
How many questions each email can ask in 24 hours, counted in the log. The owner changes it with a variable, from 1 to 1000.
By construction: the refusal, unrelated questions and leaving Vesting are answered on the server before a provider is chosen, so they never call the model.
Declared by the manual: the production provider, Groq, runs on its free plan, which doesn't charge; it only cuts off when the quota runs out, and then the chat switches to local search.
