Case study — careers
Averil.
A career operating system for the person doing the work — not the recruiter hiring them.
The problem
Career advice is written for whoever is paying for it.
Almost every tool in this category earns its money from employers. That shapes everything downstream: the advice optimises for filling a role, not for the person's next five years. The rest of the market sells motivation — courses, certificates, and rewritten résumés with no read on whether any of it is what the field is actually asking for now.
So the person doing the work is left guessing at two separate jobs at once: staying employable in the role they have, and building toward the one they want. Nothing on the market treated those as one system.
The brief we wrote for ourselves: build the thing that tells someone the truth about their own market position, and then what to do about it this month.
What we built
Four surfaces, one loop. Each one answers a question the user actually asks out loud.
Standing
Where you actually are
A read on the user's current skills against what their field is hiring for — stated plainly, including the parts that are weakening. No score theatre, no gamified badge.
The gap
Between now and next
The distance between the job they hold and the one they're aiming at, broken into work that fits around a full-time role.
This month
Small, real moves
A short list of actions with a reason attached to each. If a recommendation can't be justified in one sentence, it doesn't ship.
What it won't do
Designed-in refusals
No mass-applying on the user's behalf, no employer seat, no resale of candidate data. Each of these was a real product decision with revenue attached — and each one is the reason the advice can stay honest.
How we worked
Four phases. The product only got its name at phase three — that's the lab rule.
Study
Interviews with people mid-career, plus a read of what the incumbent tools optimise for and who pays them.
Prototype
A single screen answering one question, put in front of real users before any account system existed.
Name & ship
It survived users, so it earned a name, a domain, and a roadmap. Averil went live on averil.ai.
Run it
The build team operates it. Every support thread lands with the people who wrote the feature.
Stack & architecture
Four layers, each replaceable. The evaluation layer is the one we'd never cut.
Layer — signal
What the market is asking for
Ingest and normalise role requirements over time, so "in demand" is a trend rather than a snapshot.
Layer — profile
What the user can actually do
Structured skill and history model owned by the user, editable by the user, never sold on.
Layer — reasoning
Model work, tightly scoped
The model compares, explains, and drafts. It doesn't decide anything the user can't see the reasoning for.
Layer — evaluation
Guardrails before features
Every recommendation path is tested against cases where the honest answer is "don't move yet". Advice that can't pass doesn't ship.
Named technologies per layer are deliberately left off the public write-up.
Screens
The product as it stands today, straight out of the running build.
Averil — the standing screen
Averil — the gap view
Averil — this month's moves
Outcome
What's true today, with the figures reported as they're confirmed.
In production
Averil is live at averil.ai, run day to day by the team that built it.
Held the line
Zero employer seats sold and no candidate data resold — through every revenue conversation so far.
—
People using Averil
Figure pending publication.
—
Weeks from prototype to live
Measured from phase two to phase three.
—
Retention or return rate
Measured over rolling ninety days.
Founder's note
The hardest part wasn't the model. It was agreeing, in writing, to the things Averil would refuse to do — and then not quietly walking them back when revenue asked.
What we'd do differently
Three things we got wrong first time, kept in the write-up on purpose.
Ship the refusals earlier
The "won't do" list was written after the first build, not before it. Next time it's part of the prototype brief.
Fewer surfaces at launch
Two of the early screens existed because they were easy to build, not because anyone asked for them.
Measure the advice, not the app
Engagement told us nothing useful. What matters is whether a recommendation changed something in someone's actual career.
Next case study
LobeStack
Marketing intelligence that runs the whole loop — and closes it.
Build with the lab