The forward-deployed engineer, encoded
The AI labs hire forward-deployed engineers to make AI work inside your company. UNBLOCK is that engineer, as software. It connects to the machines and agents you already run, learns what your company knows, ranks what to fix first, then starts working.
Fig. 01The first sixty seconds, compressed. Every row keeps the sentence it came from.
01The specialist bottleneck
The industry's answer to “AI doesn't work here” is to send a human.
88%
Of organizations now use AI
Regularly, in at least one business function. Adoption already happened — the bottleneck moved past access.
≈⅓
Have begun scaling it
Nearly two-thirds haven't started scaling beyond pilots. Most AI use stays experimental or local.
5%
Achieve value at scale
Across more than 1,250 firms studied. The distance between using AI and running on it is where the value sits.
80%
Of workers are out of capacity
They lack the time or energy for the work they already have — and 53% of leaders say productivity must still rise.
Everyone the specialists
can’t reach is our market.
The labs are betting on people — OpenAI’s deployment company launched with roughly 150 forward-deployed engineers on day one. Specialists can deploy AI. They cannot democratize it. Their job is to encode the operating layer, not become it.
The full argument is the company thesis, at kaeva.app
02The first shift
One download, and the work starts before anyone writes a ticket.
Fig. 02What it finds on a laptop
Reads the room. Finds the AI tools already on the machine, the repos you touch, the rhythm you work in. It reads and never writes, behind a deny-list that keeps raw paths and secret values from leaving.
Fig. 03Four spellings, one matter
Builds a model of your business. Every person, tool and project it sees becomes an entity, and every entity keeps the verbatim quote it came from. Merges are reversible, so nothing gets overwritten.
Fig. 04One list, in the order it costs you
Tells you what to fix first. Needs are scored on how often it shows up, how much it hurts, how fixable it is and how fast it pays. A zero on any one axis kills the whole score.
An agent’s own approval is ignored.
Fig. 05Where the machine stops
Does the work, and shows its receipts. Reads run on their own. Writes stop at a proposal until a person moves them, and an agent's own approval doesn't count.
What it hands back is a table with rows, and a source quote on every row. You can check its work against the sentence it read.
03One memory, every agent
This is why the second week costs less than the first.
Every decision, correction and dead end becomes a block: signed, hashed, scoped, and readable by any agent you run. New members, human or agent, inherit what the company already worked out instead of asking someone. When a person leaves, their thoughts stay.
Fig. 06One brain, read and written by both
Every block carries who wrote it and when. Humans and agents use the same identity system.
Provenance is a required field. A memory that can't name where it came from doesn't get written.
Private, team and org-wide. Writes default private; you widen them deliberately.
The graph models disagreement as a first-class edge, so the brain can show you where it disagrees with itself.
04Bring your own agents
It wires the stack you already have.
One pass writes the integration settings each app needs, and safely skips anything already configured.
Only with your permission. It shows you what it found, then waits for you to choose Add or Not now.
Reachable over MCP from any host. Bring an agent and it joins the same brain.
mcp.kaeva.app/mcp05Trust
An autonomous engineer sits on a trust boundary.
It doesn't borrow your API key. Each agent holds its own credential, chained to the human who owns it, so every action has an actor and an accountable owner.
Revoke a person and every credential belonging to the agents they own dies in the same transaction.
A proposed action stays proposed until a person moves it. The machine can propose; only a human can make it actionable, and an agent's own approval is ignored.
No conversation is read without a recorded yes. The schema rejects any other value, so nothing can start an import around you.
Secrets caught on the way in are recorded as a pattern and a placeholder, never a value.
A query tells you how much of the brain sits outside your reach, as a number and never as content, then lets you request access from whoever owns it.
We publish what survives a hostile read, and mark what is still coming. This page describes the controls in the architecture as it stands today. It does not imply a certification or an audit we have not completed.
06Operating note
UNBLOCK was built by the fleet it ships: humans and agents coordinating through the same shared memory this beta gives you. The numbers below are that brain, counted.
201 tests. Mutation-proven invariants.
Mutation-proven means deleting a guard turns a specific named test red. Checked by hand, one guard at a time. Production counts code-verified 2026‑08‑11.
07Private beta · power users · small teams · early enterprise
Free during the private beta. The desktop app and the MCP endpoint are live today.
Questions before you install? hello@kaeva.app
Turn the AI tools already on your computer into one system, without becoming your own integration engineer.
Give teammates the same guided setup and the same shared memory, one computer at a time.
Start with an invited group, learn on real work, and expand only when the beta earns it.