AI Hiring Platform

Talentos

Every profile in the pool read against the job description, checked for fraud, and argued over by a jury of AI agents before anyone reaches a shortlist.

TypeAI hiring platform

SurfaceWeb app, desktop-first

RoleProduct design, UI system, frontend

Year2026

Talentos pipeline dashboard with pool metrics, open requisitions and shortlist — The pipeline view — pool size, qualified count, open requisitions, and what's aging out.

The brief

A single requisition draws a pool no recruiter can read end to end. So the pool gets filtered by keyword — which rewards whoever writes the best résumé rather than whoever can do the job, and cannot tell a padded claim from a real one.

Talentos reads the whole pool instead. Every profile is scored against the job description, checked for planted keywords and fabricated history, and weighted by what a person has demonstrably spent time doing rather than what they listed under Skills. What survives goes to a panel of AI agents that argue the case before returning a verdict a human can read.

Score the whole pool, not a sample

Every profile is read against the requisition, scored, and banded into tiers. The distribution is published rather than hidden — most of a real pool sits low, and showing that is what makes the rest credible.

Catch the honeypots

Planted keywords and impossible timelines surface as integrity flags, kept separate from the fit score. A strong-looking profile can still be held back, and the reason is written down.

Weight what they actually did

Recency-weighted time per area, derived from work history, sits beside the claimed skill list. The bar labelled search & ranking measures the work, not the keyword.

Let the agents argue

Finalists go to a jury of AI agents. Each returns plain-language reasoning split into what's good and what to watch, so the verdict can be audited instead of trusted.

Keep the pipeline honest

Stalls surface as warnings against their SLA — stuck in screening for eighteen days — so nobody quietly ages out of a process nobody is watching.

Design language

Seven rules do most of the work. They exist because the product shows a great deal of data at once and still has to be readable on the fifth hour of a screening session — every one of them is a decision about attention, not taste.

Monochrome by default; colour always means something

The interface is one near-black ink on off-white, every neutral a tint of the same value. Hue appears in four places only: unread counts, SLA breaches, live status, and the leading score.

Why

The product's job is to put thousands of numbers on one screen. If the chrome is colourful, nothing reads as urgent. Reserving hue for state makes a flag impossible to miss and makes the palette self-documenting — colour here is a data type, not decoration.

Exactly one inverted card per row

In every metric strip a single tile is black-on-white; the rest are grey-on-white.

Why

Density needs an anchor. Inversion sets reading order without a second typeface, a larger size, or an accent colour — and it can move to whichever metric matters on that screen, which a fixed size hierarchy can't do.

Panels are fills, not boxes

Cards are flat light-grey surfaces with a generous radius and no border. Structure comes from the gaps between them.

Why

At this density, borders build a grid of cages and every line competes with the data inside it. A fill separates regions at a glance and keeps the screen quiet enough to work in for an hour at a time.

Bars instead of charts

Every distribution — locations, experience, tiers, momentum, skill recency — is the same row: label, full-width track, fill, right-aligned value.

Why

One shape learned once, then read everywhere. Identical track lengths make separate panels comparable to each other, and the right-aligned number keeps a sliver legible when the bar itself has nothing to show.

Headings ask questions; the interface answers in prose

Sections are titled What's good, What to watch, What they've actually worked on, Is their activity rising or falling? The summary is a sentence about a person before it is ever a number.

Why

The output is a judgement about someone's career. Plain language keeps the model auditable and stops a score carried to three decimals from feeling more objective than it is.

One dark surface, and it never moves

A narrow black rail is pinned to the left edge. Every other surface in the product is white or light grey.

Why

Content changes completely between screens; navigation is the only constant. Making it the single dark element means orientation costs nothing and the workspace never competes with the chrome.

Numbers carry the typographic weight

Metrics are set several steps larger than any heading near them, with their labels small and quiet alongside.

Why

Nobody opens this product to read headings. The type scale should agree with why someone is here.

Screens

Every surface in the product, in the order a recruiter meets them.

A candidate summarised in a sentence before any number appears.

Talentos candidate evidence view with recency-weighted skill areas and history — Recency-weighted evidence: what the work history shows, next to what was claimed.

Talentos analytics with score distribution curve and tier distribution — The score distribution, published rather than hidden.

Talentos analytics breakdown by location, experience, title and momentum — Where the qualified pool actually sits — by city, experience, title, and momentum.

Where it stands

Screening stops being a keyword query and becomes a reviewable document. Every shortlist position carries the reasoning that put it there, every hold carries the flag that caused it, and the shape of the whole pool stays visible behind both.

Next case studyLoom