Moka Eva
Eva is an AI recruiter that screens candidates before a human does. I designed the calibration system that lets recruiters teach it what they're looking for without having to write it down first. Shipped 0→1 in four months — 43% adoption, 54% day-1 retention.
AI Hiring Agent
SaaS
CONTEXT
An enterprise HR vendor needed a product that didn't feel enterprise.
Moka serves companies like Nvidia, Tesla and DJI with a full HR suite. To differentiate in a crowded market they bet on Eva, a lightweight, AI-first recruiting tool. I joined as founding product designer and shipped it in four months.
USER RESEARCH
The hard part isn't screening candidates. It's knowing what you're screening for.
Recruiters sit between candidates and hiring managers. In interviews the same three struggles came up: turning a vague request into clear criteria, judging fit with confidence, and getting through manual evaluation fast enough to matter.
INSIGHT
Recruiters can't name their criteria until they see a candidate.
Criteria aren't static. They're implicit: recruiters can't articulate what they want until a real profile is in front of them. They evolve as the review goes on. And they're multidimensional.
So Eva had two jobs: help recruiters find criteria they couldn't name yet, and help them trust a match once they saw one.
PRODUCT DECISION
If criteria can't be declared, they have to be discovered.
This reframed the product. Instead of a form asking recruiters to define requirements upfront, Eva had to be a loop: the recruiter and the AI arriving at the criteria together, one decision at a time.
The loop runs in three steps. Eva reads the job description and proposes initial criteria. It presents candidates one at a time, and every accept or reject sharpens the model. Then it asks for correction where it's still unsure.
STEP 01 · INTAKE
Guidance should arrive inside the writing, not before it.
Recruiters enter a job description and Eva extracts a structured hiring brief. Rather than have them approve AI output as a separate step, I put the assist inline, a magic wand offering support mid-sentence, so recruiters keep control of the words.
I compared tabs, forms and an editable document for the brief itself. Tabs split it across screens; forms were too rigid for nuanced criteria. The editable document won.
STEP 02 · CALIBRATION
Ten candidates at once got skimmed. One at a time got judged.
Once the brief is set, Eva starts showing candidates. I tested presenting all ten together, it overwhelmed people and they stopped reading closely. One at a time held their attention. After a run of rejections, Eva asks why, turning a dismissal into training data.
STEP 03 · REFINEMENT
Most people won't correct an AI unless it asks.
After calibration, recruiters move to the full candidate list. Waiting for them to volunteer feedback didn't work. The final design has Eva initiate, asking targeted questions in chat, so correcting it feels like answering rather than filing a complaint.
workspace
Eva had to sit beside the work, not in front of it.
Every final screen so far puts Eva on the left and the work on the right. That layout came out of testing. In a full-screen chat, every candidate meant a round trip: open the profile, read it, come back, pick the thread back up.
Four of five recruiters preferred the split view. One said it felt like pinning key candidates to a wall and glancing at them while chatting.
Evidence
Recruiters didn't want a rating. They wanted a reason.
Inside the split view, the question became how to present a candidate. Five-star scores tested badly: people wanted to know why someone fit, not how much. Expanded detail views were dense and slow. The answer was qualification highlights paired with an AI summary explaining each match.
OUTCOME
Recruiters found a match in two or three rounds instead of giving up.
Eva shipped in four months.
The calibration loop closed the gap between what recruiters meant and what the AI returned. The clearest sign: people stopped rejecting their way through the entire list.
REFLECTION
A good AI product teaches the user what they actually want.
Eva improves from feedback, but the more interesting effect ran the other way: recruiters came out of calibration understanding their own criteria better than when they went in. That's the bar I'd set now, not just that the system learns, but that the user does.
The check I'd build next is the counterweight: if recommendations start skewing toward one demographic, the product should say so before a recruiter has to notice.
