Openjobs AI - MIRA：Your AI HR Agent
Worked as the sole designer on the team to automate the recruiting process.
Role: Solo Lead Designer
Company: OpenJobs AI
Scope: User Research, Product Strategy, UX/UI Design, Prototyping, Visual Design
Timeline: Q2 - Q3 2025
Overview
As the solo lead designer, I owned the end-to-end design process — from user research and prototyping to building the design system and integrating AI and LLM capabilities into conversational and multi-agent experiences.
Problem
Recruitment today is slow, fragmented, and costly, even in tech-driven organizations.
-
Inefficient screening: HR teams spend over half their time filtering unstructured resumes.
-
High data noise: Job descriptions and CVs vary widely in format and quality.
-
Slow response cycles: Candidates wait up to a week for replies, reducing engagement.
-
High operational cost: Each hire costs thousands in manual HR hours.
Design Challenge
As the solo designer for OpenJobs AI, I faced four main design challenges:
-
Understanding Recruiting Workflows
I needed to deeply understand how founders and hiring managers recruit — from writing job descriptions to sourcing, outreach, and interviewing. Each step revealed unique pain points that shaped our product experience. -
Delivering the Most Relevant Candidates Efficiently
Using existing candidate databases, LinkedIn data, and AI-driven inference models, I had to design a system that surfaced the most relevant and high-intent candidates for each role. The goal was to help hiring managers instantly identify the best matches and spend less time screening. -
Engaging Candidates and Building Interest
A key challenge was helping early-stage companies attract top talent who were not actively job hunting. I designed automated yet personalized outreach flows — allowing companies to introduce their story and culture through AI-generated, human-like email templates that improved response rates dramatically. -
Designing Solo at Speed and Communicating Cross-Functionally
As the only designer, I owned all deliverables — from user flows, prototypes, and design systems to high-fidelity visuals. I collaborated closely with the PM and CTO, aligning on feature priorities and clearly communicating desired outcomes to engineers. This required fast iteration, strong system thinking, and clear design documentation to maintain quality and consistency.
Design Process
1. Research — Understanding the Recruiting Ecosystem
I started by interviewing founders and hiring managers from companies like Modal Labs, Dyna Robotics, Stytch, ByteDance, and Pylon.Through these conversations, I mapped the full recruiting workflow:
-
Founders and hiring leads struggled to define “must-haves” precisely.
-
Recruiters wasted time sourcing irrelevant candidates.
-
Candidates lacked awareness or motivation to explore smaller or newer companies.
I also analyzed existing tools like LinkedIn Recruiter Panel, Simplify, Jobright, JuiceBox, Mercor and HireEZ to identify gaps where AI could add unique value — such as conversational JD creation, multi-agent collaboration, and adaptive candidate ranking for the clients, our business partners since we mainly focus on B2B.
Based on my understanding, I mapped the full recruiting lifecycle to identify bottlenecks and automation opportunities.
User Research & Personas
To understand real hiring behaviors, I interviewed multiple startup founders and hiring leads to uncover their goals and pain points. Below are key personas that shaped the product experience: (I'm also a venture fellow at Eastlink Capital so I got the chance to know the founders through several events):
.png){kind=small}
Erik Bernhardsson — Founder, Modal Labs
Goal: Hire top AI infrastructure engineers with strong inference and distributed systems skills.
Pain Points:
-
Recruiters couldn’t understand technical must-haves (e.g., Databricks, PyTorch inference).
-
Many good engineers weren’t actively looking.
Design Insight: Needed a way to converse about hiring goals and surface hidden, high-fit passive candidates.
Jason Ma — Co-founder, Dyna Robotics
Goal: Scale a robotics engineering team quickly across Asia and North America.
Pain Points:
-
No in-house recruiter; needs a one-click way to contact top engineers.
-
Needed to attract top-tier engineers from companies like Boston Dynamics who didn’t know Dyna yet.
Design Insight: Built one-click outreach with personalized email templates to introduce Dyna’s vision and invite top candidates — improving engagement by 5×.
Lydia Gorham — Head of Product, Stytch
Goal: Fill product related technical roles while balancing daily management duties.
Pain Points:
-
Too much time spent reviewing irrelevant resumes.
-
Recruiters misunderstood nuanced product–engineering crossover skills.
Design Insight: Designed a “feedback loop” UI to mark candidates as “match / not a match and why,” improving AI matching precision over time.
Fangfei Chen — Data Center Lead, ByteDance
Goal: Identify specialized AI infra engineers for global data centers.
Pain Points: Difficult to filter by rare skillsets (e.g., inference optimization, GPU scheduling).
Insight: Added structured must-have filters and AI-powered skill inference from public GitHub and LinkedIn data.
.jpg){kind=small}
Tom Zhou — Engineering, Pylon
Goal: Hire cost-effective new grads while prioritizing those with real industry internship experience (so they ramp faster and require less senior hand-holding).
Pain Points:
-
Resume volume is high and noisy: many applicants are fresh but lack practical experience.
-
Hard to surface internship-quality signals (internship project depth, mentor references, real product contributions) from standard profiles.
Design Insight: Prioritize filters and scoring for internship experience (GitHub repos tied to internships, project write-ups, internship duration and mentor/company reputation). Add UI affordances like an “Industry-Ready” badge and quick filters for internship signals to help Tom find the best entry-level hires fast.
2. Ideation & AI Integration — Designing Multi-Agent Collaboration
To design an AI-native solution, I mapped the end-to-end recruiting lifecycle through interviews and research:
-
Job Definition: Hiring managers often express requirements verbally or in messy job descriptions, making it hard for recruiters to find matches.
-
Candidate Sourcing & Matching: Recruiters filter manually through LinkedIn, resumes, and proprietary datasets. Time-consuming, prone to errors.
-
Candidate Engagement: Candidates may not know or trust the company; engagement is low without personalization.
-
Decision & Feedback: Managers need ways to approve or reject candidates while training the system to improve future recommendations.
This workflow mapping revealed pain points and opportunities for AI intervention, especially around structured JD creation, candidate ranking, and engagement.
I focused on how AI could augment recruiting workflows without overcomplicating the product. Although we often talk about “agents” conceptually (Job Briefing, Search, Engage, Monitor), in practice I envisioned a unified AI system that evolves through conversation.
3. Wireframing & Prototyping
After mapping the full hiring manager journey, I translated the my thoughts into concrete user and screen flows to visualize how each step connects — from defining a new role to engaging the right candidates.
3.1 User Flow and Screen Flow (Hiring Manager Journey)
I started by outlining the end-to-end hiring manager flow:
Login → Create Role (MIRA chat) → Review AI Matches → Give Feedback → Send Outreach → Monitor Results.
Each stage translated into a primary screen or micro-interaction:
-
Dashboard: Overview of all active roles and candidate pipelines.
-
MIRA Conversational Setup: Dialogue interface for defining hiring needs.
-
Candidate Match List: Top recommended profiles with explainability cards.
-
Candidate Detail View: AI reasoning panel showing skill overlaps and provenance.
-
Outreach Composer: Editable, AI-generated message templates.
-
Analytics / Feedback Screen: Visualizes performance metrics and candidate intent signals.
These screens formed the core system map, showing where AI augments each decision rather than replaces it.
3.2 Rapid Iteration and Low-Fidelity Wireframes
I usually began with low-fidelity sketches to validate information hierarchy and conversational pacing:
-
Ensured that MIRA’s chat interface always exposed both dialogue and structure preview, allowing users to see the evolving JD in real time.
-
Designed a “summary drawer” that compiles all confirmed details into a living job description.
After validating flow efficiency with target users, I moved to high-fidelity Figma prototypes, focusing on clarity, trust, and system responsiveness:
-
Integrated AI reasoning cards and confidence indicators in the candidate list.
-
Added microinteractions (loading pulses, typing indicators) to humanize the AI experience.
-
Built a consistent design system (colors, tokens, agent components) for scalability and rapid iteration.
3.3 Key Features
1. Onboarding / Start a New Search
-
Trigger: User logs in or clicks “Start New Role.”
-
System Goal: Understand the hiring context.
-
Interaction Type: Conversational (MIRA agent).
Dialogue Design — From Conversation to Structure:
The most critical piece of prototyping was designing the dialogue flow between the Hiring Manager and the AI Agent (MIRA).
The goal: turn informal, natural speech into a structured and comprehensive job description.
Example dialogue:
MIRA (AI Agent): “Hi Jason, let’s define the hardware engineer role you want to hire for at Dyna Robotics. What’s the main goal for this position?”
Jason Ma (Co-founder, Dyna Robotics): “We need a senior hardware engineer who can design and deploy high-performance robotic manipulators for autonomous mobile platforms. The role is on-site in Redwood City.”
MIRA: “Understood. Could you describe the essential technical skills — the must-haves?”
Jason: “They must have deep experience in electromechanical design, actuator/motor selection, sensor integration, and embedded control systems. We also need someone who’s built production-grade robotic platforms before — ideally with experience in motion control and hardware/software integration.”
MIRA: “Are there particular companies where you’d expect this experience to come from?”
Jason: “Yes — engineers who’ve worked at Boston Dynamics or other leading robotics companies in the Bay Area would be great. For example, competitors like Figure AI, Vicarious, Agility Robotics, or Bright Machines — they all have the kind of hands-on robotics R&D and system integration background we’re looking for.”
MIRA: “Got it. Any preferred technical stack or additional qualifications?”
Jason: “Yes, ROS/ROS2, CAN/EtherCAT communication protocols, and mechanical system calibration are critical. Ideally, they’ve taken a prototype from concept to field deployment.”
MIRA: “Thanks. Here’s what I’ve summarized so far:
-
Role: Senior Hardware Engineer – Robotic Manipulation & Mobile Platforms (onsite in Redwood City)
-
Must-have skills: electromechanical design, actuator/motor selection, sensor integration, embedded hardware/software integration, ROS/ROS2, CAN/EtherCAT.
-
Preferred background: engineers from Boston Dynamics, Figure AI, Vicarious, Agility Robotics, Bright Machines, or similar Bay Area robotics startups.
Would you like to adjust or expand any of these points?”
Jason: “No, that’s perfect. Let’s generate the detailed job description and share it with potential candidates.”
MIRA: “Excellent. I’ll draft the full JD and prepare an initial candidate list based on these criteria.”
➡️ Output: A structured, machine-readable Job Profile (role, skills, experience level, location, company preferences, etc.)
2. AI Candidate Search & Matching
-
System uses:
-
LinkedIn public data
-
Open-sourced personal websites (e.g., GitHub, Kaggle)
-
Private datasets from AGI House and OpenJobs AI’s network
-
-
Goal: Deliver 3 best-fit initial candidates.
Flow:
-
System analyzes structured JD → retrieves potential matches.
-
Presents summary:
“We found 3 initial matches. Please review and tell us who fits best — and why.” -
User reviews cards for:
-
Experience overview
-
Key skill alignment (highlighted keywords)
-
Prior companies / open-to-work signals
-
Confidence score
-
Decision Point:
-
👍 Mark as “Good Fit”
-
👎 Mark as “Not Fit”
-
✍️ Add feedback (e.g., “Too junior,” “Wrong stack,” “Great inference work”).
➡️ Feedback is stored → AI retrains ranking algorithm → improves next round of matches.
3. Candidate Shortlist & Insights
-
Goal: Give hiring manager an overview of top-ranked, high-intent candidates.
Screen Elements:
-
Candidate list view with tags like “High Intent,” “Industry-Ready,” “Past Databricks.”
-
Hover to view AI confidence reason:
-
“Matched based on AI inference paper co-authored at Stanford (2022).”
-
-
Sort and filter by skill depth, company background, location, or intent signal.
➡️ User selects candidates to contact.
4. Outreach (AI-Generated Email Templates)
-
Goal: Help hiring manager or founder contact passive candidates efficiently.
Flow:
-
User clicks “Contact Selected Candidates.”
-
System drafts personalized outreach templates using:
-
Company context (e.g., “Dyna Robotics just raised $120M Series A”)
-
Candidate’s expertise (e.g., “Your recent Databricks work aligns perfectly with our inference optimization challenges.”)
-
-
User can:
-
Preview email
-
Edit tone or add personal message
-
Send directly or schedule
-
➡️ With one click, outreach emails are sent.
➡️ Candidate replies trigger automated interview scheduling suggestions.
5. Feedback, Continuous Learning & Dashboards
-
Conversion rate dashboard shows hiring manager their recruiting efficiency.
-
Hiring manager feedback (e.g., “Focus more on Databricks alumni” or “Avoid PhDs”) feeds back into MIRA’s model.
-
24/7 behavioral monitoring, auto-prioritizes high intent candidates.
➡️ System continuously refines candidate matching and outreach tone.
6. AI Touchpoints Throughout
3.4 Prototyping Deliverables
-
Full interactive Figma prototype for the hiring journey.
-
Conversation flow map visualizing all AI-human dialogue states and edge cases.
-
Design system library including reusable components for agent cards, chat bubbles, feedback inputs, and explainability panels.
-
Developer handoff specs with annotated flows, motion guidelines, and API mappings for conversational data exchange.
4. Testing & Iteration
After finalizing high-fidelity prototypes, I ran iterative usability tests with pilot clients.
Testing Goals
-
Measure speed, clarity, and trust in AI interactions.
-
Validate whether hiring managers could complete key tasks without guidance.
Method
-
8 participants (Modal, Dyna, Stytch, Pylon, Bytedance).
-
Moderated sessions + Figma prototype.
-
Scenarios: JD creation, match review, outreach.
With these key findings, I refined our prototype and hand over to the CTO, engineering and PM for MVP development.
Results & Future Steps
I had 8 pilot sessions across Modal Labs, Dyna Robotics, Stytch, Pylon and ByteDance.
1. Quantitative Outcomes
2. Qualitative Insights
From Founders & Hiring Leads
-
“It finally feels like the AI understands how I think about hiring.” — Tom Zhou (Hiring Manager, Pylon)
-
“The outreach messages felt like they were written by us — candidates actually replied.” — Jason Ma (Dyna Robotics)
From Recruiters
-
Found AI explanations critical: “Seeing the reason behind each match turned AI from a black box into a partner.”
-
Appreciated feedback control: “I can teach the system instead of restarting every search.”
From Candidates
-
Responded more positively to context-aware outreach (“I saw your funding round — sounds exciting!”).
-
Fewer drop-offs thanks to human-sounding tone and personalized role highlights.
3. Business & Team Impact
-
Early-stage companies without recruiting staff could now handle end-to-end hiring autonomously.
-
Reduced recruiter hours saved an estimated $80 k per year for a 10-person startup.
-
Demonstrated cross-functional collaboration efficiency — design, product, and AI engineering aligned around one structured workflow.
-
Strengthened partnerships with data sources (AGI House, LinkedIn API) by proving measurable value of their datasets.
4. Future Product - Level Iterations
