Applied research and engineering
Get one AI workflow
working in your world.
A task that loses its context. An automation that fails quietly. An agent that says “done” without the result. Bring one concrete problem. We agree the outcome, make a bounded change, and leave you the checks and handover.
Describe your workflowYour existing tools can stay. No platform migration required.
The first engagement
Working AI
Workflow Pilot
$1,500 for an agreed milestone
One workflow in one test environment, using up to two supported interfaces. We confirm feasibility, access, scope and the delivery date before payment.
Target: five business days after usable access, with up to twelve delivery hours. We only accept work that fits the agreed limit.
What you receive
- A clear starting point and observable acceptance criteria.
- One implemented change or a clearly labeled working prototype.
- Checks you can rerun, including relevant failure and recovery cases.
- Source or configuration, a plain-language runbook and one written handover exchange.
- One same-scope correction window within seven days.
Our agents are not independent third-party certification. Acceptance depends on the result and the checks agreed with you.
Not sure what the repair involves?
A separate $300 feasibility milestone covers up to three working hours. You receive a workflow/data map, the demonstrated constraint, an acceptance plan and a build quote or an explicit reason not to proceed. The fee is credited toward a subsequently agreed $1,500 pilot.
Evidence before assurance
“The run ended”
did not mean
“the lesson helped.”
Our own agent project contained a false-credit path: a stopped child could make a lesson look helpful without a recorded work outcome.
We replayed the same five checks on original and corrected source snapshots. The original failed four subtests; the corrected version passed all five methods, while the contrasting valid cases stayed intact.
This is a project self-audit using fixture evidence, not a customer result or proof of generally reliable learning.
Read the example and its limits ↗The company behind the work
A particular result.
A larger program.
Jessy Brenenstahl is the maker of Ultimentality and founder of RQ MCP LLC. The program spans formal and executable semantics, agent memory and orchestration, desktop and phone systems, and interactive fiction.
Its core runs today: Relicquary (v0.9.0), durable, local-first, receipt-bound memory and orchestration for agents. One Rust binary, SQLite and Markdown on your disk, a receipt for every mutation, and your model, local or your own provider. AIM wraps that core in governance; it is a working prototype, and we mark that line rather than blur it. A result only counts when it survives a test it could have failed.
These engagements apply relevant parts of that work to a bounded customer problem. They do not promise a generally self-improving agent. Engineering may be AI-assisted; any agreed human review is performed and recorded.
Start with the problem
What needs to work?
Email the workflow, desired result, current tools, budget and deadline. Redacted examples or a test environment help us establish fit.
hello@relicquary.com ↗We confirm fit and terms before payment. No open-ended support or whole-system rebuild is included.