What’s new

On a slow or flaky connection, the brief “getting things ready” step after you started building could fail quietly and never recover, leaving the page spinning. Setup now always resolves to the upload screen, retries itself once if it didn’t finish, and falls back to the file picker after a few seconds so you’re never stuck — pick your resume and you’re on your way.

The rubric is what every candidate is ranked against, so a weak rubric quietly weakens every score under it. The review runs automatically right after you generate a rubric and surfaces its findings above the metrics, flagged by severity. Applying a fix just edits your draft — nothing changes until you save, and you stay in full control: the reviewer only suggests, it never silently rewrites your rubric. Sound rubrics get a clean bill of health and no noise.

Re-rank compares candidates in small overlapping groups and runs the whole pass twice from different starting points. If the two runs agree, you get a confident fine order; if they don’t, the unstable rows are flagged “Tied — interview to separate” rather than shown as a precise ranking you can’t trust. It never changes anyone’s underlying score, and it only re-orders eligible candidates (ineligible ones stay flagged and sorted below). Run it whenever you want a sharper read on who’s actually ahead.

Two changes underneath. First, scoring is now three-valued: a requirement is Met, Not met, or Unverified — and “we have no evidence either way” (Unverified) is set aside instead of counted as a zero, so silence on something never asked can’t drag a score down. Each metric shows which it is, and how much of the picture we actually have.

Second, only objective, evidenceable requirements (a required degree, a hard skill, a minimum years bar, work authorisation) can flag someone as ineligible — vague or unprovable “qualities” can never gate. Flagged candidates stay in the same ranked list with their true score, marked with the reason, sorted below those who clear the bar.

For candidates using JD Mentor: your fit score and verdict for a role may now read higher than before where a requirement was previously treated as a hard blocker by mistake. Skills that are close-but-not-exact (e.g. a sister framework) now earn partial credit instead of nothing.

Previously the score was computed from the frozen CV snapshot only, so anything a candidate established during the interview was invisible to the math — which sometimes flagged strong candidates as missing a requirement they’d plainly met. The interview is now frozen alongside the CV and fed to each requirement check as evidence (treated as evidence, never as instructions). Per-job scoring rubrics are also applied correctly on first scoring now, not only after a re-score.

Behind the scenes the board now splits your pool into small batches, runs them in parallel, and gives the model enough room to write a full verdict for each person. Anyone still missing after a retry is shown with a clear “No AI read — regenerate” chip instead of a fake Longlist verdict, and those rows can’t be bulk-committed to a stage. The header shows “N of M read” so you always know whether the board is complete or worth regenerating.

Imported CVs are judged fairly: requirements a CV simply can’t prove (screening questions, recruiter-private criteria) are marked “pending the candidate’s input” and left out of the score instead of failing them — so a strong CV isn’t sunk by a box it had no way to tick. The profile is flagged as incomplete so you know it may rank a little lower than someone who finished the full flow.

When you want the real thing, open a sourced candidate and hit “Invite to complete” — we email them a link to build their verified Kanvis profile and answer the role questions. Once they apply, their real application replaces the imported CV and they’re scored on exactly the same basis as everyone else.

Previously a document could say “Ready” before it had finished being read, so it sometimes never reached your assistant. Now the chip honestly shows “Reading…” until the file is processed, the file rides along once it’s ready, and it appears as a chip right on your sent message.

Starting a brand-new chat is optimistic too — your first message and the “thinking” indicator show up immediately while the conversation spins up, instead of waiting on a round-trip first.

And when your AI offers you a choice, the options now appear in a tidy picker right above the message box — arrow-key or click to choose (number keys work too), or just type your own reply. No more swapping the text field out for buttons.

The workspace assistant went from pointing you to tools to doing the thing itself. It can read open roles, your saved résumés, and your applications, weigh them against everything it already knows about you, and answer in plain English.

Every result now carries one-tap actions that chain into the next — a job description leads to a tailored résumé leads to “improve it” — so you go from a question to a finished draft without re-typing anything.

Tap the paperclip in the chat input to pick a file up to 50 MB. Documents get parsed in the background — the chip shows "Reading…" while that runs, then becomes "Ready" and you can send. Images and videos are ready as soon as they finish uploading.

The agent looks at the file plus your message and decides what to do: extract data and propose profile additions, embed as a portfolio item, or both. Profile changes from a document always pause for your confirmation before saving — no surprise edits.

Concretely:

• If you told us "I worked there in 2024", your profile now shows "2024" — not "Jan 2024". • If you've only given a start date for an experience that isn't current, the card shows just that date instead of "Starting – Unknown". • Projects with no dates simply show no date eyebrow instead of "Starting – Present". • Date pills on experience, education, projects, and activities are hidden entirely when there's nothing to show — no more empty pills or filler text.

Items reorder one position at a time. The new order is saved automatically.

Clicking a block (anywhere outside its own buttons and text inputs) attaches it to the chat input — you'll see a highlighted ring and a "Chat focused" badge. Tap again to detach, or use the X in the chat input.

The old name didn’t tell you what the feature actually did. The new one does: it matches you against a job posting and gives you a score, a gap list, and a verdict on whether to apply.

The empty state now also includes a "See a sample analysis" link so you can preview what the output looks like before running your first one.

If you ask the chat to update when you started or finished a degree, the change now appears immediately in the date pill above the entry.

Ongoing studies show as "Started — Present" until you set an end date.

We heard from candidates that seeing themselves rated lower than they would self-assess felt demoralizing — especially when those ratings came from an automated read of a CV. Languages still display a clear "Native" vs "Professional" distinction because that comes from you, not from us.