Kitsuno is not a job board. It is a pipeline — a sequence of specialized AI agents that turn raw job postings into scored, validated, draft-ready opportunities. This article explains how each stage works, what decisions are made, and where the human stays in control.
The pipeline: six agents in sequence
Every job posting that enters Kitsuno passes through six stages. Each stage has a different agent, a different AI model, and a different purpose.
Stage 1 — Scanner. Crawls job sources on a schedule. The frequency depends on the plan: once a day for Kit (free), up to four times a day for Pro. Sources include global platforms, national employment agencies, niche boards, and ATS platforms via X-Ray search. The Scanner pulls raw postings and stores them for processing.
Stage 2 — Extractor. Reads each raw posting and extracts structured data: job title, company, location, employment type, seniority signals, language requirements, salary (when available), and a geo-scope classification (worldwide, country-restricted, or city-restricted). This structured data powers everything downstream.
Stage 3 — Geo-scope gate. Before spending resources on full scoring, the system checks whether the job is geographically possible. It compares the extracted location against the user’s work permit regions. A job restricted to a country where you have no work authorization is deleted before scoring. This gate also handles remote ambiguity — “remote” alone does not mean international unless the posting explicitly says so.
Stage 4 — Scorer. The core intelligence step. An LLM receives the full job description and the candidate’s structured career evidence (skills, experience, education, certifications — pulled from the Professional Record Store). It evaluates fit across four rubrics and returns a structured score with specific matches and gaps.
Stage 5 — Writer. When the user requests a draft, the Writer agent assembles context from the career library, the scoring data, and any existing Fit Report, then generates a tailored CV, cover letter, and application email. Three sequential AI calls produce each document section. The output uses the user’s configured voice, language, and emphasis settings.
Stage 6 — Validator. Every AI-generated document passes through 14 automated quality checks before the user sees it. Issues are flagged visibly. Hard failures offer a free re-draft. The validation results feed back into future drafts through a self-correction loop.
The four scoring rubrics
Every job is scored on a 100-point scale across four weighted dimensions. The weights are configurable per search profile — these are the defaults.
Role alignment (40 points). How well do the job’s requirements match your demonstrated skills and experience? The scorer looks at responsibilities, domain expertise, and technical requirements against your Library evidence. Not keyword matching — semantic understanding of what the role needs and what your career demonstrates.
Location and remote compatibility (25 points). Does the geography work? The scorer factors in your work permit regions, preferred locations, accepted work modes (remote, hybrid, onsite), and the job’s specific location requirements. A hard disqualifier applies: if the job is restricted to a region where you have no work authorization, the location score drops to near zero regardless of how good the role fit is.
Seniority fit (20 points). Is the experience level right? The scorer evaluates leadership scope, years of experience signals, and management expectations against your career history. You can configure preferred seniority levels and exclude levels that are not relevant.
Organization type (15 points). Does the employer type match your preferences? Corporate, startup, NGO, government, academic — different environments suit different people. The scorer evaluates the posting’s organizational signals against your stated preferences.
The total score produces a fit verdict: strong fit, moderate fit, weak fit, or not a fit. Every score shows its breakdown — you can see exactly which rubric contributed what, which skills matched, and which gaps the AI identified.
The validator: 14 checks, zero tolerance for fabrication
The Validator is a dedicated agent whose only job is catching the other agents’ mistakes. It runs on every document before delivery.
Hard failures block delivery and offer a free re-draft. These include: empty content, template placeholder artifacts left in the text, AI self-reference leaks, language mismatches (requested German, got English), contact information leaking into cover letters, metric hallucination (numbers or percentages that do not exist in the career library), and first-person violations.
Warnings are shown as readable flags the user can review and dismiss. These include: unusual document length, overused filler words (leverage, synergy, utilize — the verbal equivalent of an empty suit), company and role name inconsistencies across documents, and an LLM-powered check for invented claims.
The self-correction loop. Validation results are stored after every draft. When a re-draft is requested, the previous failures are injected into the Writer’s context as mandatory avoidance instructions. The AI knows what went wrong last time and actively avoids repeating it. Draft, validate, store failures, re-draft with awareness, validate again.
The Validator is calibrated, not omniscient. Small models have structural ceilings. The design principle is readable, dismissable flags — not zero false positives. Transparency about limitations is itself a quality signal.
Multi-model architecture: the right model for each task
Kitsuno does not use one model for everything. Different tasks have different requirements — speed, accuracy, creativity, cost — and different models serve them best.
Scoring and extraction use models optimized for structured JSON output. The task is classification: given this job and this career, produce a structured evaluation. Smaller, faster models handle this reliably because the output format is constrained.
Writing uses larger models optimized for natural language generation. Cover letters and professional summaries require nuance, voice consistency, and the ability to synthesize evidence into compelling narrative. Paid users get premium models for this task because the quality difference is visible.
Validation uses fast, lightweight models. The task is binary classification (pass/fail) on specific checks. Speed matters because validation runs after every draft.
Concierge chat uses models balanced for conversational quality and context depth. Paid users get deeper context windows and more capable models, but the personality and voice stay identical across tiers.
EU-first routing. Personal data from paid accounts is routed exclusively to EU-based AI providers under normal operation. This is not a marketing claim — it is an architecture constraint enforced in the routing layer. The specific providers, models, and fallback chains are documented, tested, and monitored.
Free-tier cascade. Free users get the same scoring quality through a cascade of free-tier providers. If one provider is down, the system falls back to the next. The scoring rubrics, weights, and evaluation criteria are identical — what changes is the provider, not the methodology.
The brain: deterministic intelligence
Not everything requires an LLM. Kitsuno’s Brain is a pure SQL and Python diagnostic engine that runs 21 rules against the user’s data — zero AI cost, zero hallucination risk.
The Brain checks things like: library completeness (how many skills have supporting evidence), pipeline health (application-to-response ratios), source effectiveness (which sources produce strong matches), profile configuration quality (missing weights, empty keyword lists), and readiness signals (is the profile ready for meaningful job searching).
The Brain generates a knowledge document for each user that feeds into the concierge chat context. When Kitso says something specific about your career, it is often the Brain talking: deterministic observations drawn from real data, not LLM speculation.
What you can verify
Every stage of the pipeline produces inspectable output.
The Scorer shows its four-rubric breakdown, key matches, and gaps for every job. The Writer produces documents you review before sending. The Validator shows its check results as quality notes — green, amber, or red — with specific flag descriptions. The Brain’s diagnostics feed Kitso’s observations, which name specific jobs, scores, and skills.
The pipeline is not a black box with a number at the end. It is a sequence of transparent steps, each with visible inputs and outputs, where the human decides what happens next.
Explore the open-source EU job source directory on GitHub →
For the principles behind these technical decisions, read Why Kitsuno exists. For live market data from this pipeline, visit Market Pulse.