Evidence-based profile / vacancy intelligence / product building
HoiJob
What started as a personal investigation into my actual skills grew intoHoiJob: a live job board and Control Room that structures vacancy texts, connects profile evidence and keeps matching controllable.
The case in 30 seconds
HoiJob
- Problem
- My CV did not sufficiently represent my actual capabilities, while vacancy texts made it difficult to compare fairly due to marketing language.
- Solution
- An evidence-based profile linked to a modular route for extraction, normalization, taxonomy, matching, scoring and human review.
- Role and focus
- Product architecture, AI-native direction, audit and job UX
- Status
- Vacancy bank live · product development active
- Result
- Live vacancy bank, Control Room and Operations Back Office · 2,715 unique vacancies substantively processed · 2,793 recorded audit runs.
Live lab / try it yourself
From text to insight
Search the live job board and see how source information, job detail, and extraction combine to create a more readable foundation for the next step.
The personal reason
My CV said too little
My resume mainly showed job titles. It showed much less clearly which processes, projects and responsibilities I had actually carried over the years.
Hospitality, sales and gaming activities formed parts of my career, but they did not automatically reveal that I also design systems, improve processes, organize projects independently and can bring complex issues from strategy to implementation. As a result, I was often judged by the label of a job, not by the capability behind it.
I didn't want to make that a better story than the facts allowed. I wanted exactly the opposite: to honestly investigate what I could demonstrably be able to do, how much experience there was behind it and which positions matched that. Do not apply above my level, but also do not end up below my level again because a CV provided too little context.
Don't make it bigger
Making actual skills visible
Do not judge smaller
First map myself
From claim to evidence
TheStefanJob Search System started as a factual inventory of experiences, skills, years of experience, strength of evidence and interconnections. Every important claim had to be traced back to work, projects, periods or concrete results.
Evidence profile · reference date 2026
Fifteen skills with professional capacity
A targeted selection based on relevance and demonstrable length of experience. The size is not the claim; the value lies in the complete self-examination and the choice of which capabilities can really support my next step.
No.
Professional skill
Domain
Experience
01Demonstrate leadership in teams and operationsPeople management24+ years
02Manage and deliver projects independentlyProject & product23+ years
03Analyze and resolve operational problemsOperations22+ years
04Conduct negotiations and close dealsSales & strategie21+ years
05Operational management and implementationOperations21+ years
06Have operational control and ultimate responsibilityOperations20+ years
07Ensure operational preparation and readinessOperations20+ years
08Analyze and improve operational processesProcess optimization19+ years
09Monitor and adjust KPIs and performanceStrategie & analyse18+ years
10Structuring and optimizing workflowProcess optimization18+ years
11Drawing up business casesStrategie & analyse17+ years
12Making decisions under pressure and time constraintsOperational control16+ years
13Align collaboration between teams and disciplinesPeople management13+ years
14Guiding change within teams and processesChange management11+ years
15Analyzing and structuring complex systemsSystems thinking10+ years
The first system
Grown from the spreadsheet
With the profile as a basis, I started loading vacancies into Google Sheets and investigating with Apps Script: which position really requires what I can demonstrably do, where is the connection strong and where is evidence lacking?
That prototype proved the core, but each new source, formulation, and reprocessing increased the interdependencies. When Sheets and Apps Script could no longer reliably support the amount of records, relationships, and exceptions, I opted against more patches. I had the system redesigned as a modular web product with separate responsibilities.
From source to responsible step inHoiJob
- 01Profile proofSkills, experiences and connections
- 02Matching prototypeGoogle Sheets and Apps Script
- 03Modular platformOwn data and processing layers
- 04HoiJobReview before action in Control Room and public site
Product architecture
One system, clear layers
Profile and vacancy only come together after source, extraction, normalization, taxonomy and review have each held their own responsibility.
Operational evidence
From organization to follow-up
Three detailed demo examples show how business context, a field signal and access per task come together in the same operational layer.
Curated demo data from built workflows. This is not a claim about third-party market adoption, a full helpdesk platform, or a certified enterprise access layer.
Learn to read vacancies
The gist of the text
Vacancy texts are rarely uniform. Hard requirements, soft preferences, cultural language and marketing are intertwined. That's why I taught the system not only to search for words, but to recognize information within three fixed questions.
What are the requirements?
Education, experience, skills, language, certificates and other conditions.
What are the tasks?
The work that someone will actually perform and carry.
What is the offer?
Salary, conditions, development, working method and concrete compensation.
Eval-driven development
Quality became a feedback loop
2.715unique vacancies are processed substantively
2.793recorded audit runs
Historical MySQL count · reference date September 1, 2026
- 01 · inputVacancy text
- 02 · extractionRequirements, tasks, offer
- 03 · auditResult
correct?yes: publication · no: recovery - endpointPublication-ready
- 04 · error patternRecord deviation
- 05 · improvementSharpen logic
- 06 · retestReprocess
A strict training requirement is different from a desired level or a preferred formulation.
Concrete responsibility had to be separated from general marketing and employer language.
Salary, conditions and development only count if the source actually provides information about them.
New spellings and missing source information had to remain visible for review.
Technical substantiation
Under the hood
The visual above is the quality loop. This includes a limited processing chain: capturing source, qualifying text, deterministically structuring, checking with AI and also returning human field signals to system improvement after publication.
-
Source with provenance
Search profiles send allowed provider routes. Each candidate holds provider, source link, source context, and a deduplication trail; there is no login orLinkedInscraping route.
provider · url · raw hash -
Not every lead is allowed to participate
An audit requires sufficiently complete, genuine vacancy text. Short snippets, consent walls, error pages, collection pages, truncated material and AI-built source sections are blocked.
full text · structure · source status -
Rules before interpretation
PHP modules clean up the text, remove platform noise, create a section map, and separate requirements, tasks, offerings, and risk signals. Boundary and validation rules monitor the categories.
clean · sections · output snapshot -
Source vs output
The AI does not rewrite the vacancy. It compares the real source with the extraction result and reports only detectable misses, noise, truncation, or duplicates in a strict JSON schema.
source quote · output reference · score -
The outcome remains testable
Extraction snapshot, model, prompt version, score, error types, times and any review needs are stored per audit. Additional guards re-check the AI result for known noise.
mysql · model · prompt version -
Recovery is not an automatic action
An error pattern — also from live human use — becomes a targeted line or code change with regression testing; this is followed by reprocessing and re-audit. A clean audit makes it ready for publication, but Operations and human field control continue to monitor the live outcome.
repair · regression · published
Human field control
Ready for publication is an intermediate status
Internal control and 'Report problem' may generate a new error signal after publication. This is not just solved for one vacancy, but feeds the quality loop that improves the entire workflow.
- 01Field signal
- 02Error pattern
- 03System restore
- 04Reaudit
Example: A task point cut off prematurely. Do not manually smooth out one vacancy, but make the cause reproducible, targeted repairs and reassess the relevant set. The quality loop remains necessary as sources, formats and languages continue to vary.
AI under frameworks
The AI assesses; limits the system
The AI here is not a free text generator or final decision maker. He compares source and extraction along defined questions and only returns a verifiable judgment.
- 01Decision question
Is this a requirement, task, offering or noise?
- 02Proof required
Each deviation links source quote to output.
- 03Guards test along
Schedule, score and validators limit the outcome.
No silent self-training. Quality increases because prompt versions, decision criteria, quality gates and regression tests are tightened in a targeted manner. So a nuance can be — for example HBO/WO in IT instead of alone HBO/WO — as a testable rule back in the audit.
Open technical review
What the iterations prove
HoiJobdoes not improve through one free model, but through a fixed method: point out deviation, tighten the rule and reassess the relevant set.
Source and output remain linked
The live demo shows that vacancy texts are already cleaned and separated into requirements, tasks and offers. Source quotes, output references and fixed guards keep every AI judgment testable.
An error becomes system knowledge
A concrete error pattern becomes a rule or code change, regression test and reaudit. In this way, one observation is returned as a targeted improvement to the relevant set.
2,700+ checks
Processing has already gone through real source and sector variation. The method proves this, without pretending that all future variants are already covered.
The biggest lesson fromHoiJob
Not feasibility, but scale
When data, relationships, and exceptions became larger than a spreadsheet process could support, scale became a concrete product choice.
-
01 · START
Spreadsheet over
Google Sheets and Apps Script were the right base for exploration. Source status, queues, errors, reprocessing and audit trails required their own chain.
-
02 · NOW
A bounded system
Source, text quality, extraction, AI audit, recovery and publication are separate steps. This means that every boundary is visible, testable and can be improved in a targeted manner.
-
03 · NEXT PHASE
Capacity follows ambition
More volume requires worker capacity, resource caches, and measurable quality control: the same chain, on a larger scale.
Why vacancy text asks so much
Every vacancy combines nuances
Requirements and tasks vary simultaneously with education level, field, certifications, software, language, industry jargon and local formulations. Offer is often recognizable; the real variation is in the meaning and combination of requirements and work.
The opportunity lies in the choice of scale. HoiJobis already working on real vacancies. For robust Dutch coverage, approximately ten times more varied data is the next step; international languages then require their own examples, rules and validation. The route is ready for organizations that see product value in this.
Scope of this loopThis is the working intake, extraction and audit chain. Publication-ready is the verifiable outcome at that time; robustness grows with field signals across more and more source and language variation. Normalization, taxonomy, matching and scoring remain separate modules; they do not constitute automatic human judgment.
From technical system to product
That's whereHoiJobwas born
The original matching tool grew into a product with its own name, public vacancy experience and protected Control Room.
Profile data, extraction, modules, database, APIs, build status, training and audit come together in the development environment.
Vacancies, companies, promotions, imports, publication, support and reports are given an operational work route.
The job board makes offers searchable and presents source, content and extraction without hiding the original text.
profilevacancyextractionreviewpublication
Built product scope
More than a vacancy list
The Control Room is built around the components needed to not only display a vacancy product, but also to manage, control and further develop it.
Candidate context
Preferences, experiences, skills, capability links, matching, scoring and review.
Manage offers
Jobs, discovery, import and sync, companies, promotions and publication status.
Making a platform work
Customer support, marketing, SEO, pages, reports and product modules.
Manage responsibly
Finance, user roles, access, database, APIs, security, training and audit.
One product · three work surfaces
From building to using
The same content chain moves through three environments: in development the system is formed, in operations the offer is managed and on the public site the result becomes usable for candidates.
Transparent validation framework
What works, what has been proven
HoiJobis a capability proof with working product layers. The case does not claim any external market validation that is not yet there.
Search, filters, vacancy details, source information and readable presentation.
Profile management, vacancy workflow, review, publication status and operational modules.
2,715 unique vacancies substantively processed through completed audits; 2,793 audit records make repeat audits visible.
Guiding and verifiable, not intended as automatic human judgment.
A commercially applicable product base, without claims about customers, turnover or proven market traction.
The original mission remained standing
A fairer next step
HoiJobstarted with my need not to be judged bigger or smaller than my demonstrable experience warranted. That personal reason became the product promise: look beyond a job title, make visible what a job really requires and keep the final decision human.
A match is therefore not an automatic judgement. It is a well-founded direction for a better conversation, with a view of what is appropriate, what is still uncertain and what evidence can support the next step.
Next cases
More systems thinking
From AI workflows through to community, interface work and entrepreneurship.
Voice agents / business operations
AVENLO
Four real-time voice agents, four brand worlds, and one bilingual phone experience with bounded business actions.
Premium yacht concept / 3D prototype
AURELIA 67
An independently constructed yacht concept study: a design exploration in which creative direction, 3D, film and interaction form one premium product story.
Thailand / property intelligence
DOMORA
An internationally applicable, multilingual real estate toolset that is the first validation market in Thailand to connect listings, local housing requirements, map search, matching, seller intake and social content.
Conceptcase / merk & product
BaitFever
Prelaunch brand and commerce experience for anglers; product lines, storytelling and the commercial market route are further expanded in a controlled manner.