Talentera
Hiring software · CV Search · designed and built end to end
CV Search in Talentera. A recruiter thinks in sentences, so I replaced the keyword box, the four match modes and the 23 filters with one composer. You write or say who you need, in English or Arabic. It underlines what it understood, shows every filter as a tag, opens the full filter set inside itself, and folds to one line once the results are on screen.
A real page in the product, at /dashboard/candidates/search. Below is the same box rebuilt for this page, with invented people. It is live, so type in it.
In 60 seconds
01
The problem
Recruiters had to translate how they think into dozens of filters: four match modes, then 23 filters in five categories, to say one sentence they already had in their head.
02
The insight
Recruiters don't think in database queries. They describe the candidate they want.
03
The design
Turn natural language into a structured, editable search: one composer that underlines what it understood and shows every filter as a tag, in English or Arabic, typed or spoken. The full filter set opens inside it, and on results it folds to one line.
04
The challenge
The AI had to be transparent and controllable, never a black box: every filter it sets can be seen, changed or removed, so it never guesses in silence.
23 → 1
Filters in the specification, now one box
Specification vs built
2 states
One composer: open to write, folded to one line on results
Built
2 scripts
English and Arabic, typed or spoken, one reading
Built, covered by tests
The outcome. No adoption figures are claimed: this is designed and built end to end, rebuilt here with invented people.
01Where I started
The page before: a keyword box, four modes, a Filters menu, and tips.
The filter wall
23 filters in five categories
Personal information
Experience
Target job
Education
Additional information
The thought, and the form
A recruiter thinks: “a bilingual medical secretary in Doha who can schedule appointments.”
Four panels opened and five boxes ticked, to say a sentence they already had.
02What I built
Underlined words are what it understood, and each became a tag. Press an underlined word to change it, open Add filters for the full set, press the mic to speak. Then search: the composer folds to one line and the cards take the page. Click it to open it again.
Describe the person or type what you know.
Write it, say it, or pick an example.
03The second pass
The second pass came from watching it used. Every filter reads as words and as a tag, the full filter set opens inside the composer, and once the results are on screen the composer folds away so the cards own the page. Each piece became a design-system component, so the next surface gets it free.
One composer, two states
Open while you write: the sentence, its tags, scope, mode and filters. Folded on results to its wording alone, so the cards own the page. Focus it and it opens again. The send button is one object that travels with the bar, never two that swap.
Filters inside the composer
Segmented tabs, a dashed cell grid measured to the width so no column is left alone, and a count beside every value. A pick is written into the sentence: tick Doha and the box reads “in Doha”.
Every filter is a tag
An icon for its kind and a round flag for places. Five show and the rest wait behind +N, so the composer stays tidy. Filters with no word form, such as age or degree, are tags too, so nothing is hidden.
One record number, one row of marks
The CV ID comes first, as one component across the product. Folders, job pipelines and duplicates became a quiet row of icon marks, with numbers beside them and the detail on hover. Sanad is an outline chip with its assessment on hover.
04What I held myself to
Nurse in Riyadh
01
Show your working
A box that guesses in silence is one a recruiter stops trusting. It underlines what it understood and shows the filters as chips, each removable.
Bilingual medical secretary in Doha
3 people in your database match this right now
02
Teach by showing
Nobody reads instructions for a search box, so the empty state types a sentence and shows what it becomes. The number at the end comes from the real engine.
03
Every count is a price
A number beside a filter is who you would have left if you picked it. A value that would leave nobody is not offered.
Reem Al-Harbi
Registered Nurse · Riyadh
04
Same person, same card
A result is the talent pool's own candidate card. The search adds the match score and the reasons, never a second design for the same person.
05
Arabic and English, both first class
One box, both scripts, one reading. A name written in one script matches the same name in the other.
Matched on
06
Say why
Every match can say why it matched, so a recruiter checks instead of trusts. A mark is a reading of fit, never a judgement of the person.
05Say it
Recruiters describe people out loud all day, and in Arabic speech removes the keyboard-switching entirely. A spoken sentence gets the same reading, underlines and chips as a typed one.
First try: a slashed mic
It read as muted, the opposite of what is happening.
Now: a stop square and a ring
A tinted button and a ring that breathes out. Listening reads as listening.
06Around the box
The full set opens as a panel inside the composer, to narrow what the sentence did in the first 80%. Every count shows what choosing it would leave, and every pick is written back into the sentence. The screens around it each got a designed answer.
Browse before you ask
Four clusters on the start screen, each bar a search, flags where it is a country.
Saved and recent, with alerts
Each with an empty state of its own, because both are empty on day one.
Compare two to four people
A comparison grid extracted from the Duplicate drawer, so one component serves both.
Confidential, on purpose
Masked until revealed. Revealing asks first and is logged with a name and a time. The log has no delete.
Add a candidate from the +
One quick step, so a modal, not a drawer. Hover the + and it says what it does. The old search-by-CV-link row is gone.
Zero results offer a move
Not an apology: the one filter to drop that finds the most people.
07Decisions worth arguing with
Any of within a group, and between groups
A person lives in one city, so Riyadh AND Doha matches nobody. The specification says the opposite. I flagged the conflict rather than silently following it.
A rule-based reader, not a model
Deliberate for a concept: it is predictable and testable. The chips are the contract, so a model can replace the reader later without changing the design.
The match mode, demoted
Every word, any word, exact phrase or boolean is a quiet menu in the toolbar that only touches the keyword part: a refinement after the sentence, never a gate before it. Where to search stays visible, because it changes what a result is.
A clustered start instead of an empty one
Someone with no query still has somewhere to start.
+Under the hood
For the designer or engineer who wants to keep going: the reading rules, Arabic and English, spoken numbers, the components and the honest edges.
08What I took from it
A recruiter thinks in sentences, not filters. So CV Search is one box that shows its reading and never hides it, and the filters became the second step.