JobEmber. AI Job Application Agent
Duration
March 2026-Present
Product Owner, Product Designer, Vibecoder
JobEmber (Own Startup)
Me (Product Owner, Designer ), Denys H (Developer, Security Expert), Claude Code
AI-Product, B2C, Startup
Problem
How might we remove the mechanical burden of job searching so people can spend their time on the decisions only they can make?
Solution
Agents search platforms, profile employers, and produce application documents that sound like you wrote them, because the generation is constrained by samples of your own writing. A detection pass runs on every output to catch anything that reads as machine-made before it reaches a recruiter. Everything flows through a five-stage dashboard; a Chrome extension pulls in positions you find while browsing. We tested it with real users across a closed beta and consistently saw application time drop by 87.5%, with testers completing eight applications in the time they'd previously spend on one. The product is multi-tenant, production-ready, and currently in the branding and pre-launch phase.
How I did the Research
Market
Before writing a line of code, I ran a market research pass. Tools like LazyApply, Sprout, and Sonara existed, but they operated on a spray-and-pray model: auto-apply to hundreds of jobs with a generic CV, no customization, no voice matching, no company research. The niche we found was the opposite end: fewer applications, higher quality, output that actually sounds like the person sending it. That gap drove every product decision that followed.
Sub-agents
It started as three agent scripts running in a terminal: job-searcher, employer-profiler, and cv-adapter — each a Claude Code skill triggered manually, outputting raw text, no interface. That was enough to validate the pipeline. Once it was clear the approach worked, the web interface followed. 20+ agents run today, now triggered from a five-stage app, but the terminal output is still what's actually happening underneath: discovering jobs, researching employers, tailoring CVs and cover letters, scoring relevance, checking for AI detection risk, tracking applications through the full pipeline. Multi-tenant, production stack, tested with real users across a closed beta.
UX decisions and why I made them
Pipeline system
The first web version ran on a dark interface, built fast to get all five stages visible in one product: discover, research, generate, export, apply. Testers flagged WCAG contrast issues in this version immediately — one of the first things we fixed in the redesign. What it proved was the architecture. Each stage working, each one handing off to the next.
UX rebuilt
UI redesign
The warm cream palette came from the interviews with job seekers. They kept returning to the word "exhausting." Job platforms trend cold. We went the other way, informed by research on emotional design in high-stress contexts showing warm palettes reduce perceived cognitive load.
The UI redesign started in Claude Design, not Figma. I'd describe screens; Claude generated and iterated on them. Figma handled the design system: components, tokens, spacing, but ideation happened in conversation. This pass added pages and features that were missing entirely.
Voice matching
Early CV versions were technically correct but didn't sound like the people using them, testers flagged this immediately. When a hiring manager reads a cover letter and then speaks to the candidate in a screening call, a voice mismatch creates doubt. Hard to name, easy to notice. We built a voice capture layer: users paste samples of their own writing — emails, previous cover letters, anything they have. The system analyzes patterns: sentence length, vocabulary level, formality, rhythm. Those become constraints on generation. The CV sounds like the person because it's derived from how they actually write.
Past Applications
Once testers started submitting real applications through the system, they needed somewhere to track what happened. Every submission is stored with the exact CV and cover letter version that was sent. Before this existed, testers were keeping separate spreadsheets or losing track of what they'd sent to whom. Filter pills were added after watching how people actually searched their history, let you slice by stage, company, or date in one click.
Experience
The original product had a one-shot Evaluator: input a project idea or a course, get a BUILD/SKIP verdict from an AI agent. The verdicts were often good. But testers who were actively job searching weren't trying to decide what to learn next — the feature pulled attention sideways. We removed it. What replaced it is an Experience tracker: users maintain a record of projects and education: what they built, their role, outcome, dates, tools. That record feeds CV generation directly. Portfolio-based hiring has been growing across product, design, and tech roles; structured, user-verified data is more actionable than a one-time verdict.
UI/UX redesign
via Claude Code
Implementation ran through Claude Code. Position color-coding by geography made the job list scannable at a glance. A template selector let users choose between Standard and ATS document output per application. Sorting and filtering gave the applications tracker real utility. A price calculation surface made the cost per action visible — relevant once the product moved toward a paid tier. Each feature came directly from something testers couldn't do without.
Chrome extension
The agent pipeline finds jobs across platforms automatically, but testers kept finding good roles while browsing that the scraper missed — niche boards, company career pages, LinkedIn. The extension closes that gap. One click on any posting pulls the position into the pipeline: title, company, description, requirements, salary, apply URL. It also reads the application form in real time, so Q&A fields are pre-populated from the user's profile by the time they reach Stage 5. Both versions shown — dark and warm — connect to the same API. Same pipeline, same data, triggered from the browser instead of the search agent.
User Testing
Throughout the build, three to five people used the product for real job applications, not test scenarios. They were in it from early on, which meant feedback arrived as actual problems: a cover letter that didn't sound right, applications they'd lost track of, a feature that pulled their attention away from what they were trying to do. Each design iteration was shaped by what the previous version broke or missed. The filter pills, the geography color-coding, the voice settings, the decision to cut the Evaluator — none of that came from assumptions. It came from watching how people actually used the thing.
Outcome
Across testers, application time dropped from around two hours to fifteen minutes — 87.5% less time, eight times faster. The time came from the research and formatting layers: the parts that don't need judgment, just attention. The fifteen minutes that remained was the part that actually needed a person.
Denys H. led security architecture and backend infrastructure throughout the build. Branding and marketing are the current phase, the product launches soon.