How Kariyan's AI Actually Works

"AI" has become a marketing word. Every job platform claims to use it; few explain what the AI actually does. This post is the explanation. We will walk through the four stages of Kariyan's pipeline, in plain English, and call out exactly where machine learning adds value, where it does not, and where you (the candidate) stay firmly in control.

Stage 1: Career Page Discovery

Kariyan does not scrape job boards. We crawl company career pages directly, a Telekom listing comes from telekom.de/karriere, a Zalando listing comes from corporate.zalando.com, and so on. There are roughly 7,500 active career pages in our index right now, refreshed on a schedule that ranges from hourly for high-volume employers to weekly for smaller companies.

This stage uses almost no AI. It is mostly a deterministic crawler: fetch the page, follow the "Open Positions" link, parse the HTML structure, identify individual job postings. We use small classifier models for two narrow tasks: detecting whether a page is a job listing or marketing copy, and extracting the structured fields (title, location, posted date) when the page has no schema.org markup.

Why career pages and not LinkedIn?

Career pages are the source of truth. By the time a job appears on Indeed or LinkedIn, it has already been on the company's own page for hours or days, and the most motivated applicants have already submitted. Kariyan reads from the source, so you see fresh roles before they reach the crowded platforms.

Stage 2: Job Extraction

Once a posting is captured, an extraction model converts the free-form HTML into structured fields: required skills, preferred skills, seniority hint, work model (remote/hybrid/onsite), language requirements, salary range if disclosed, and a short summary.

This is where the first non-trivial AI shows up. We use a fine-tuned LLM with a constrained JSON output schema, and a deterministic post-processor that validates every extracted field. If the model emits a salary that is not numeric, the post-processor drops the field rather than guess. The data you see in your matches is what we could verify, never what we hallucinated.

Stage 3: Profile Scoring

For every new role, we compute a match score between 0 and 100 against your profile. The scoring function is deliberately simple and explainable:

• Skill overlap against your verified skills (from your CV and the voice onboarding)
• Seniority alignment based on years of experience and role keywords
• Location fit against your preferred cities or remote preference
• Language match between the posting and your fluency levels
• Disqualifiers like "5+ years required" when you have 2

Each match comes with a "why this score" explanation listing the matched signals and the misses. We do not use a black-box neural ranker because we want the score to be defensible, if a 70% match looks wrong to you, you can see exactly which signal contributed.

Stage 4: Document Generation

For roles you choose to apply to, Kariyan generates a tailored CV and cover letter. This is the stage that uses the most LLM compute, and it is also the stage where we are strictest about what the AI is allowed to do.

The generator cannot invent facts. It can rephrase your existing bullets to emphasise the most relevant projects, choose which experience to lead with, and adapt the tone to the company's house style. It cannot add a job you never had, a degree you never earned, or a skill you never claimed. The system prompt enforces this hard, and a separate validation pass rejects any output that introduces facts not present in your profile.

The cover letter is structured to address what the specific posting cares about. If the job emphasises distributed systems, your relevant projects move to the front of the letter. If it emphasises team leadership, your management experience leads. The result is something a recruiter can plausibly believe you wrote yourself in 30 minutes , because the underlying facts are yours, only the framing is automated.

Where Humans Stay in the Loop

Kariyan never sends an application without your approval. Every prepared package lands in your dashboard with a side-by-side preview of the CV, cover letter, and email body. You can edit, regenerate, or skip. Most users approve in under two minutes; the rest spend longer because they are thinking carefully about which roles to actually pursue, not because they are fixing AI mistakes.

The Autopilot mode lets you delegate the approval step too, but only after you have set explicit rules (minimum match score, salary floor, excluded companies). Autopilot is a delegation of judgement, not of control. Currently in closed beta with a small group. Outside the beta, nothing is sent without your approval.

What We Do Not Do

We do not auto-fill application forms with fabricated work history. We do not write cover letters that lie about your experience. We do not use AI to fake responses to recruiters or schedule interviews on your behalf. The reason is simple: those things would burn your credibility the first time a hiring manager noticed, and they are not what gets you hired anyway.

Kariyan is built on the bet that the bottleneck in modern job hunting is not candidate quality, it is candidate visibility. The pipeline above exists to get you in front of the right roles fast, with materials that accurately represent what you bring. The hiring decision stays where it belongs: between you and the company.