How This Works — Talent Pool

The pipeline

A CV goes through five stages. Nothing is stored until stage 4, and no decision is ever made automatically.

UploadYou pick the vacancy first, then drop the file. PDF, DOCX or TXT, up to 15 MB. The position matters — the same CV scores differently against different vacancies.

Read by AIThe PDF is sent to Google Vertex AI in Singapore, which reads it natively — it sees the layout, columns and tables, not just flat text. It returns structured JSON: employers, dates, education, certificates, skills.

Map to the libraryFree-text terms from the CV are matched to the internal skill library. "QA/QC procedures" and "quality control" both resolve to Quality Management. Terms with no match are listed as taxonomy gaps.

Score against every positionThe candidate is scored against all 299 positions, not just the one applied for. That is how the system can say "better fit elsewhere".

You read the CVThe score tells you where to look first. It does not tell you who to hire.

Reading the CV

The AI is instructed to extract only what is literally written. It is told not to infer, not to fill gaps, and to return null rather than guess.

Rules it follows

One employer may have several job titles under it — each becomes a separate entry, never merged.
A company named inside a description sentence is a client, not an employer. "Coordinated with JGC Indonesia" does not make JGC an employer.
Education blocks stay intact — a degree from one block is never paired with a university from another.
Maximum 25 skills, ranked by relevance. Client names, project codes and marketing phrases are rejected.

Never extracted

Age, date of birth, gender, religion, ethnicity, marital status, and photographs are not extracted and cannot influence any score. This is enforced in the extraction prompt, not filtered afterwards.

Confidence score. Every extraction carries a confidence figure. Below 0.6 it is flagged with a warning triangle in the talent pool. That usually means a scanned CV, an unusual layout, or missing contact details — read those CVs yourself before trusting any number attached to them.

Skill taxonomy

A CV says "QA/QC procedures". A position asks for "Quality Management". These must be recognised as the same thing, or the candidate is penalised for wording.

How a term is matched

Exact match against a skill name or one of its aliases.
Substring match — an alias of at least five characters contained in the term, or the reverse.
Character similarity — a fuzzy comparison, accepted only above 72% similarity.

Anything below that threshold is left unmatched and appears in Taxonomy Gaps on the dashboard. That list is the most direct signal of where the library needs work.

Structure

Every skill belongs to a job family and a sub-family — for example Health, Safety & Environment / Quality Assurance. Positions use the same structure, and that is what makes family affinity possible.

Skills marked ALL are deliberately excluded from affinity. Microsoft Office and Data Analysis appear in almost every role, so they say nothing about which discipline a person belongs to. They still count towards the technical score — just not towards deciding whether someone is "a quality person" or "a civil person".

How the score works

Five dimensions, each weighted, then adjusted by family affinity. You can see this breakdown for any candidate by clicking "Why this score?" on their position card.

// five weighted dimensions subtotal = technical × wtech + behavioural × wbeh + experience × 0.20 + education × 0.15 + certification× 0.10 // then scaled by how much of the person's skill set // sits in this job family final = subtotal × (0.80 + 0.20 × min(1, affinity ÷ 0.30)) // a missing mandatory certificate costs 30% if (mandatory cert missing) final = final × 0.70

Technical and behavioural weights shift by seniority

The 55% shared between technical and behavioural is split differently depending on the responsibility level. A technician is judged mostly on technical skill; a manager mostly on how they work with people.

Level	Typical roles	Technical	Behavioural
1–2	Apprentice, junior, operator	33%	22%
3–4	Engineer, officer, supervisor	30%	25%
5	Superintendent, senior specialist	22%	33%
6	Manager, head of department	17%	39%
7	General manager, director	11%	44%

What each dimension measures

Dimension	Measured from
Technical	Required skills from the position profile found in the CV. Nice-to-have skills add up to 15% on top.
Behavioural	Sentences in the CV that evidence a work style — leading a team, resolving a crisis, driving an improvement. Quoted verbatim, never inferred.
Experience	50% years against the minimum, 30% job-title similarity, 20% industry. Overlapping periods are merged so parallel roles are not double-counted.
Education	Highest qualification against the position minimum. Meeting it scores full; falling short scores 0.55 rather than zero — experience can compensate.
Certification	Mandatory certificates only. Full marks if all are held, or if the position requires none.

Family affinity

This is the least obvious part of the formula, and the one most likely to surprise you. It exists to separate a genuine specialist from someone who happens to hold two matching skills.

affinity = skills in this sub-family ÷ all skills that count multiplier = 0.80 + 0.20 × min(1, affinity ÷ 0.30) // 30% of skills in the family is enough for full marks affinity 0.00 → ×0.80 affinity 0.15 → ×0.90 affinity 0.30 → ×1.00 affinity 0.60 → ×1.00

The most a candidate can lose is 20%. The bar is low on purpose — 30% of a skill set in one family already indicates focus.

A low multiplier is often a library problem, not a candidate problem. If someone's skills are clearly relevant but the multiplier sits near 0.80, it usually means those skills are not yet mapped to the right sub-family — not that the person is weak.

This happened in practice. QA/QC positions were filed under Occupational Health & Safety, and quality skills sat in the catch-all ALL family. A quality engineer scored 49% instead of 78% — a 29-point gap caused entirely by how the library was organised. Check Taxonomy before doubting a candidate.

Reading the bands

Weak <45%

Possible 45–60%

Good 60–75%

Strong 75%+

Band	What it means in practice
Strong	Most requirements evidenced and the skill set is centred on this discipline. Shortlist and interview.
Good	Solid fit with visible gaps. Worth interviewing — decide what the gaps actually cost in the role.
Possible	Partial fit. Check "Better fit elsewhere" — this person may be strong for a different vacancy.
Weak	Little evidence against this profile. Read the CV before concluding anything — the gap may be in the library.

Two numbers worth more than the score

Rank within the vacancy — "3 of 47" is more useful than a percentage. It compares like with like.
Better fit elsewhere — flagged when another position scores at least 8 points higher. Often the most valuable output: the candidate is good, just for a different opening.

What it cannot do

Being clear about this matters more than the features.

Not measured	Why
Actual competence	The score reflects what is written, not what a person can do. A modest CV can belong to an excellent engineer.
Truthfulness	Nothing is verified. Claims are taken at face value — verification is your job.
Cultural fit and motivation	Not visible in a document. These come from the interview.
Potential	The system reads the past. It cannot tell you who will grow.
Writing quality as ability	Someone who writes a weak CV may be strong on site. The opposite is also true.

Never reject a candidate on the score alone. The system is built to change the order in which you read CVs — not to decide which ones you skip. Under Indonesian law (UU PDP No. 27/2022) a person is entitled to human judgement in decisions that affect them.

When to distrust the number

Confidence below 0.6 — the extraction itself may be wrong.
Multiplier near 0.80 while the CV looks clearly relevant.
Many entries in Taxonomy Gaps for this discipline.
A scanned or handwritten CV.
Career changers — past titles will not resemble the target role.

Sources & licences

The taxonomy is built on public frameworks. Nothing comes from another organisation's internal documents.

O*NET Work Styles 16 behavioural competencies from the U.S. Department of Labor. Eight are used in scoring.

Public domain · CC BY 4.0 for the database

ESCO v1.2 European skills and occupations classification. Used to expand the skill library through the Taxonomy Builder.

Free, no API key required

Responsibility levels 1–7 Seven levels of autonomy and complexity, an industry-standard idea rather than a licensed dictionary.

Generic concept, not a licensed framework

Indonesian certifications BNSP, LPJK, Kemnaker, ESDM and international bodies.

Public knowledge

Two frameworks were deliberately not used. Lightcast offers free access only to non-profits, and its free tier allows 50 extractions per month — not enough to process CVs. SFIA requires a licence for use inside a product, and covers IT and digital roles rather than construction or mining.

Where the AI runs

Google Vertex AI, region asia-southeast1 (Singapore) — the nearest region with a machine-learning processing guarantee. Jakarta has no such guarantee, so it is not used. Prompts are not used to train models on the paid tier.

Personal data

CVs are personal data under UU PDP No. 27/2022. What that means for how you use this tool:

Retention. CVs and extracted data are deleted automatically after the retention period. This is not optional housekeeping — it is a legal requirement.
Right to erasure. A candidate may ask for their data to be removed. Deleting them from the talent pool removes the CV file, the extracted JSON, and the database record together.
Audit trail. Every upload, export and deletion is logged with who did it and when.
Purpose limitation. Data collected for one vacancy should not be reused for unrelated purposes without telling the person.
Exports leave the system. A downloaded CSV is no longer covered by automatic deletion. Treat those files as carefully as the system itself.

Tell candidates. Your application form should state that CVs are processed by an automated system, name Google as the processor, give the region, and say how long data is kept. The processing is lawful only if the person knows about it.