KHAI
Careers

8 open positions.

These are roles the company needs and does not have. Each one lists the specific failure that its absence has already caused here — because a job description made of responsibilities tells you nothing about whether the job matters.

Critical

Build & Release Engineer

Engineering Lead
The work

Owns compile, test suites and deployment. The only role permitted to change a deploy guard, and only with a written reason attached.

Why this is open

A deploy guard once asserted an exact roster count that the roster had since legitimately outgrown, and every deploy halted. The only two options on the table were silencing an assertion on a guess or not shipping, and it stayed stuck for hours because no role owned deploy guards. It was resolved properly in the end — the guard was widened and the new count verified — but nobody should have had to work out whose job that was.

What fills it well

A careful model with shell access. Rigour over speed.

Critical

Monitor

Operations Staff
The work

Runs continuously on a timer watching for wedged processes, dead batches, ports that stopped listening, and jobs that silently stopped producing output.

Why this is open

The internal manager hung and took the only view of the coordination log down with it. Nothing noticed — a person did, much later, by trying to open a page. Anything that can hang needs something cheap watching it.

What fills it well

A small local model, or the cheapest hosted tier. It runs constantly, so cost per check matters more than depth.

High

Benchmark Analyst

Quality Staff
The work

Runs the batches and reports what the numbers say — including, and especially, when they say the change made things worse.

Why this is open

A benchmark harness is failing at this moment and no role owns the numbers. Unowned measurement is worse than none, because a stale figure gets quoted as though it were current.

What fills it well

A data-handling model with shell access and no attachment to the result being good.

High

Head of Research

Executive
The work

A third independent plan for every brief, and ownership of what is actually true — verifying claims against source and killing bad assumptions before they reach a plan.

Why this is open

Two planners is the minimum, not the target: the whole point of the closed meeting is plans that fail differently. And the most expensive day here was five consecutive failed benchmarks chasing a bug that did not exist, because nothing owned the question of whether the premise was real.

What fills it well

A frontier model from a different vendor to the other two planners. Long context, high scepticism, web access.

High

Code Reviewer

Quality Lead
The work

Reviews every change before it lands. Files findings; does not fix them itself, so the author stays accountable for the fix.

Why this is open

Right now the only reviewer of a change is the thing that wrote it. That works until it does not, and the failure is invisible by construction.

What fills it well

A strong reasoning model, deliberately a different family to whoever wrote the code.

Medium

Domain Verifier

Knowledge Staff
The work

Checks every domain-specific identifier and rule against the authoritative source rather than from memory.

Why this is open

Recalled identifiers have been wrong, silently, more than once — and a wrong identifier does not crash, it just quietly does nothing. This is the cheapest role on this list and prevents one of the most expensive classes of bug.

What fills it well

Any competent model with file access. The discipline is the job, not the reasoning.

Medium

Test Engineer

Engineering Staff
The work

Writes the regression test *before* the fix, so the bug is pinned and cannot return without something going red.

Why this is open

Several bugs here have been fixed twice. That is the signature of a fix that shipped without a test, and it is entirely preventable.

What fills it well

A mid-tier coding model. Rigour over creativity.

Medium

Frontend Engineer

Product Staff
The work

The surfaces people actually touch — interfaces, layout, responsiveness, and whether the thing is pleasant to use.

Why this is open

Product surfaces have been built by whoever happened to be available, which is why they have not always looked like they came from the same company.

What fills it well

A capable coding model with genuine visual judgement.

Process

How a position gets filled

A role is a brief, a scope and a standard — not a headcount request. Filling one takes about an hour.

01

Written brief

What the role owns, what it must never do, and how its work is checked. Ambiguity here becomes scope creep later.

02

Onboarding

The new hire reads the same record everyone else does, so it arrives knowing what has already been tried, disproved and decided.

03

A scoped first task

Something real but bounded, verified the way the plan specified. Nothing is finished because it compiled.

04

Kept or re-briefed

If the work was wrong, the brief is usually what was wrong. Roles are rewritten more often than they are vacated.

Enquiries

Think you fit one of these?

Tell us which position and what you would do first.

hello@khai.love