End-User ATS Platform
I designed and built an AI-native job-application platform end to end. It sources roles across three boards, scores them against a custom fit model, tailors a verified resume for each, and tracks every application on one glanceable board. Working solo with an agentic AI harness, I combined product design, workflow automation and AI governance into a fast, fully human-controlled system.
Role
Product Designer & Solo Builder
Industry
Career / AI tooling
Team
Solo operator + AI pair
Timeline
1 week, 2026
Problem
Applying broadly without applying badly
Modern job hunting is high-volume and low-signal. Mass-applying with a generic resume is noisy and demoralising, and bespoke applications don't scale. With a hard three-week runway, I had to apply both broadly and well, without burning out or shipping low-quality work.
The stakes were direct. Application quality drives callbacks, but quality is exactly what volume kills. Several constraints sharpened the problem: a solo operator, an ADHD working style where cognitive load is a first-order limit, job boards that actively block scraping, and the catastrophic cost of a single hallucinated figure reaching a real resume. Solving it well meant working across product, automation and AI at once.
Process
An AI funnel with a human gate
I built the system on one principle: a single source of truth. SQLite is the spine and the spreadsheet a generated view, so the board and the export never drift. Sourcing is multi-modal, pulling roles over unattended requests where boards allow it and a logged-in, tethered browser where they block scraping, then a content fingerprint auto-dedupes re-listings into one card.
At the core is an agentic harness for resume tailoring: a fan-out of AI subagents drafts in parallel, then every output passes an integrity verifier that catches invented employers, migrated figures or title drift before it advances. Nothing reaches a real application without clearing that gate, and I approve every send. This is AI used with governance, not on faith.
And I designed for my own constraint. Treating cognitive load as the core requirement rather than an edge case, I built a glanceable-first Kanban cockpit where a card shows just enough to decide and everything else is one click away.
Outcome
The reach of volume, the care of bespoke
The result turned an unmanageable manual process into a calm, glanceable pipeline, with broad reach, per-application quality, and a clear next action at every moment.
The automation gave reach: 140+ roles sourced, deduped and fit-scored across three sources. The harness gave throughput with integrity: ten resumes tailored and verified in a single parallel run of about two to three minutes. And the governance gave control: zero applications sent autonomously, every one cleared by a human gate.
Together they delivered the reach of mass-applying with the care of bespoke work, without the cognitive overload that usually forces a trade-off between the two.
140+
Roles sourced, deduped & scored across 3 sources
Minutes
To tailor and integrity-verify a batch of resumes
0
Hallucinated figures shipped to a real employer
Reflection
What I'd do differently
The hardest part wasn't building features. It was learning not to trust AI output. The answer was governance, not faith: a verifier gate, change-notes, and a manual send step. That principle, let AI draft and make a human approve what matters, is the most transferable thing I took away. The real limitation was deduplication. Company names vary and aggregators hide the employer, so I shipped a pragmatic fingerprint and a reversible cleanup rather than chase a perfect matcher. If I did it again, I'd invest in that data contract on day one, before duplicates piled up. Prevention over cleanup.
End-User ATS Platform
AI Product Design
AI Automation
Harness Engineering
Product Strategy