Portfolio · AI product design · New narrative draft

JobNova: giving people control over AI job applications

As founding designer, I shaped a connected job-search experience where AI explains a match, helps revise a resume, and applies within boundaries the job seeker chooses.

Company
Liba Space
Year
2025 – ongoing
Role
Founding Designer
Partners
Product · ML · full-stack engineering
Scope
Research → AI interaction → delivery
Visit JobNova
  • 0-to-1 product strategy
  • Human–AI interaction
  • Design systems

At a glance

The problem

Matching and generation still left people judging, checking, submitting, and tracking each application.

My contribution

Connect the journey and define what AI explains, what people review, and when automation may act.

Impact

2,700+ users enabled Auto Apply, about 1,300 of them paying. Onboarding completion rose from 73% to 96% after a one-week A/B test.

01 / My role

I owned the decisions between a recommendation and an action.

JobNova helps job seekers evaluate roles, tailor materials, and manage applications in one place. I worked with product and engineering from discovery through interaction design, the UI system, and release review.

JobNova recommended jobs with match scores, manual and automatic application states, and live Auto Apply activity
The connected workspace · Job discovery and application activity share one context. Interface numbers are demonstration content, not outcome metrics.
  • Product direction

    Use research and the application journey to shape scope with the founder.

  • AI interaction

    Define explanations, review points, permissions, and recovery states.

  • System & delivery

    Create reusable patterns and review interface and output states with engineers.

02 / Research & reframing

Faster individual tasks still left a fragmented journey.

I used 20+ interviews and journey mapping to examine where people spent effort and where they wanted AI to stop. The research synthesis pointed to three connected needs.

  • Judge relevance

    Help people understand why a role deserves attention before tailoring an application.

  • Check AI work

    Make generated changes visible so the user can verify their professional story.

  • Choose control

    Support different review needs before AI submits something in the user’s name.

I reframed the product around completing relevant applications with a level of control the user chooses.

New users in the historical 30-day cohort
263
Uploaded a resume and set preferences
179
Enabled Auto Apply
51

The same cohort leaves a concrete design question: 84 people did not finish setup. The funnel locates the gap; it does not explain why they left. Anonymized research notes and step-level drop-off data are still needed.

03 / Workflow strategy

One application journey. Four different AI responsibilities.

I connected matching, customization, application, and tracking. At each handoff, I defined the information AI supplies and the decision that remains with the person.

Recorded product walkthrough · Move from JobNova’s landing page into matched jobs, filter opportunities, and review a role. Interface counts are demonstration content.
  • 01 · Explain

    AI surfaces fit and gaps. The person decides which opportunity is worth pursuing.

  • 02 · Collaborate

    AI proposes resume changes. The person can inspect and edit the content.

  • 03 · Act within rules

    AI applies under selected criteria, materials, and approval settings.

  • 04 · Account for it

    The product exposes progress, submitted materials, and what needs attention.

This made the design problem specific: reduce repetitive work while keeping consequential actions understandable and reviewable.

04 / Explain the fit

Use a score to scan. Use evidence to decide.

I separated quick prioritization from deeper evaluation: a compact job list leads into fit details and a contextual Agent. The detail view connects strengths and gaps to a possible next step.

Job detail showing fit dimensions, resume improvement cues, and an Agent explanation alongside the role
Explainable matching · Fit dimensions and contextual reasoning sit beside the role. The score represents product matching, not a measured probability of getting hired.

The tradeoff: putting every explanation on the list adds reading before users know which role matters. I kept the scan view compact and moved the reasoning into details.

05 / Review the changes

AI can rewrite the resume. The user stays its author.

I designed two review levels: an overview of what changed, followed by section-level editing. This lets people check the output and correct their story before using it.

Recorded product walkthrough · Generate a tailored resume, inspect changes, and move into editing. The recording demonstrates the interaction; it is not a timed usability test.
Resume overview with change summaries beside the section editor and highlighted document preview
Overview → editor · Change summaries support initial review; editable sections support precise corrections. Preview content is sample copy.

The tradeoff: a finished document hides the work needed to verify it. I exposed changes and editing controls rather than treating generation as completion.

06 / Govern the action

Automation needs permissions that people can understand.

I organized Auto Apply into matching criteria, autonomy, materials, and notifications. A person can decide which roles qualify, whether approval is needed, and which documents the system may send.

Recorded product walkthrough · Set matching criteria, choose an autonomy level, configure application materials, and review submitted applications. Interface counts are demonstration content.
Auto Apply settings with match strategies, three approval modes, material controls, and notification settings
The permission model · Immediate submission, a review window, and explicit approval are different commitments. Match thresholds shown here are configuration values, not performance results.
  • Which opportunities?

    Matching strategy and job preferences bound the roles automation can pursue.

  • When can it submit?

    Approval modes make the point of human intervention explicit.

  • What can it send?

    Resume and cover-letter choices define the materials used on the person’s behalf.

The tradeoff: granular controls add setup effort. I grouped them into understandable decisions; setup completion and later rule changes are the measures needed to evaluate that balance.

07 / Follow the outcome

The system must stay legible after the person steps away.

I connected application progress, submitted material versions, and replies to the application record. Queueing, submitting, waiting for approval, and failure need different feedback and next actions.

Recorded product walkthrough · Review queued and submitted applications alongside Auto Apply activity. Interface counts are demonstration content.
JobNova application dashboard with submitted applications, conversion funnel, status filters, and live Auto Apply queue
Following execution · Application states, submitted records, and live activity stay visible together. Dashboard numbers are demonstration content, not measured project outcomes.
  • In progress

    Distinguish queued work from active submission so waiting has a clear meaning.

  • Needs a person

    Separate pending approval and manual steps from unattended execution.

  • Failed or completed

    Make recovery needs visible and preserve a record of what was sent.

08 / System & delivery

I designed a shared language for AI behavior and product UI.

I brought matching, review, and application states into reusable patterns. With ML and full-stack engineers, I reviewed output quality, UI states, and recovery paths through handoff and release QA.

JobNova design system with colors, typography, spacing, controls, and navigation components
Visual foundations · A shared component vocabulary across discovery, generation, and application.
Variants of the JobNova application card for different execution and review states
Behavior in the system · Card variants carry status, explanation, and available actions together.

Delivery evidence to add: one annotated handoff or QA example showing an issue I identified, the decision made with engineering, and the final change.

09 / Results

Auto Apply reached 2,700+ users and about 40k applications.

Users enabled Auto Apply, about 1,300 of them paying
2,700+
Applications submitted by June 2026
~40k
Interview-invitation rate, detected from email
4.1%
Onboarding completion · one-week A/B test, 100+ users
73% → 96%

I redesigned onboarding from a single screen into a step-by-step flow and shipped it after a one-week PostHog A/B test. The earlier 30-day cohort in section 02 had already located the largest drop-off: the step from a finished profile to granting automation permission.

10 / Reflection

The next iteration starts with the handoff people hesitate to complete.

The core design contribution is the relationship between explanation, review, permission, and feedback. To choose the next iteration, I would locate the largest drop-off between setup, first application, and continued use, then examine the corresponding interaction.

  • Is setup too demanding?

    Compare setup completion and abandonment by step before simplifying the controls.

  • Does review build confidence?

    Pair approval behavior and rule changes with feedback about unwanted applications.