Designing an AI job application agent from zero to one — where the hardest problem was earning the right to act on someone's behalf.

Dashboard — the daily briefing

The daily briefing — Auto-Pilot banner, stat cards, and today’s matches

Entry point

The product never opens on an empty state or a raw list. It opens on a sentence: what the agent scanned overnight, how many strong matches it found, how many applications are out, and how many interviews are active right now.

Auto-Pilot status shown as an explicit banner, not a silent default — it states plainly that it’s idle until turned on

Four stat cards: matches above bar, applications sent, interviews active, daily cap used

“Today’s matches” ranked by how much of the role the profile already covers

“Recent activity” narrates exactly what the agent has done, in plain language, with timestamps

Before the candidate decides anything, they know the current state of their search — no digging required.

Jobs — discovery with the reasoning attached

Jobs — match ring, salary, and reply-time reasoning attached to every role

Find + match

Filtering feels like any job board, but the moment a role is selected, the right panel stops being a listing and starts being an argument for or against applying: match score, estimated salary, average reply time, and how many people are already competing for it.

Match score rendered as a ring with a plain-language read (“Strong”), not just a raw number

Est. salary, average reply time, and applicant volume shown alongside the score, not buried in the description

“Job requirements vs your profile” comparison, toggled between summary and detailed view

Standard filter bar — title, location, salary, workplace — kept familiar on purpose

The candidate can judge whether a role is worth fifteen minutes of review before opening it.

Résumé — tailoring with a hard constraint

Résumé tailoring — additive-only changes against a saved match

Prepare

The single most important sentence on this screen isn’t a headline, it’s a rule: tailoring “only ever adds — nothing is removed, and nothing is added that isn’t already true of your profile.” That’s a trust boundary stated as plainly as possible.

“Tailor to a job” applies changes against a specific saved match, never generically

“Edit by instruction” is instruction-based rather than open-ended chat — “add Python to skills,” not a freeform conversation

“Your edit rules” persist across every future version, so a constraint like “keep bullets under two lines” only has to be set once

A constraint stated in plain language does more trust-building than any settings toggle could.

Apply for me — the risk dial, made literal

Auto Apply — quality floor, daily cap, and a live preview of what queues

Automate

This is the human-in-the-loop framework turned into an actual control surface. Instead of a single on/off switch, the candidate sets a match-quality floor and a daily send cap, previews exactly what would queue, and only then opts in.

Match quality floor — Excellent / Strong / Good / Stretch — with a one-line consequence for each choice

Daily cap slider, explicitly warning that a high volume in a short window reads as bot behavior to employers

“Next up” shows the real jobs that clear the bar right now, before automation is switched on

“Turn on Auto Apply” is a deliberate action, never a default state

Automation is opt-in, tunable, and previewed — never a black box the candidate has to trust blindly.

Ready to send — nothing leaves without a look

Ready to send — every prepared application waits for a human

Review

Every application the agent has finished preparing lands here first, not in an outbox. “Send all” exists for speed, but each row still opens into its own review before it can go out.

Match score shown per application, so the strongest candidates for approval are easy to spot at a glance

“Prepared” timestamp shows how long an application has been sitting, waiting on the candidate

Bulk “Send all” sits next to individual “Review & send” — speed and control offered side by side, not traded off

This is the medium-risk tier from the trust framework, built as an actual screen — prepared, never sent, until a person acts.

My applications — the spreadsheet, retired

Track / Applications — the spreadsheet, retired

Track

Every application, regardless of how it was sent, lands in one table with a clear stage and a last-update timestamp — the exact structure candidates were building by hand before.

Stage shown as a plain badge — Applied, Interview, Rejected, Confirmed — colored for a fast scan

“Last update” surfaces which threads have gone quiet and which are actually moving

Applied date kept visible so nothing sits forgotten for weeks unnoticed

Replaces the manual spreadsheet with the same at-a-glance scan, populated automatically.

Progress — reading the funnel, not just the count

Track / Progress — the funnel, week over week

Track

Application count on its own hides where a job search is actually leaking. This screen reframes the question from “how many did I send” to “where does the funnel narrow,” week over week.

Four stages as rings — applied, screened, interviewed, offers — each with its own conversion rate

Weekly funnel table shows sent versus replies, with a week-over-week delta on reply rate

Deltas colored to show improvement or decline at a glance, without needing to do the math

A healthier way to read a job search than raw volume — it points at what to adjust, not just what happened.

People — the relationships behind the applications

Track / People — recruiters and referrals as a pipeline

Track

A job search isn’t just a stack of forms — it’s a set of half-started conversations with actual people. This screen tracks recruiters and referrals the same disciplined way the pipeline tracks applications.

Each contact tied back to the specific company and role they’re connected to

Status shown honestly — Connected, Pending, Not contacted — instead of collapsing everyone into one inbox

Action button adapts to status: “Message” once connected, “Reach out” when it hasn’t started yet

Nudges the candidate toward the specific next move for each relationship, not a generic “check your inbox.”

Saved searches — agents that keep scanning

Saved searches — standing agents that keep scanning

Find, standing

A saved search here isn’t a static filter a candidate has to remember to rerun — it’s a standing agent that keeps working in the background and reports back with a match count.

Each search agent is named and scoped plainly — “Senior Product Design · Remote,” “Design Leadership · Fintech”

New-match count is the pull that brings the candidate back, instead of them needing to remember to check

“Run now” available on demand, so the agent never feels like the only way in

Turns a one-time search into ongoing coverage, without asking the candidate to babysit it.