Systems around the work.
We design the layer around your documents, workflows, and decisions. A person approves what carries risk. You keep the system when the engagement ends.
See one workflow run.
Work passes through sources and checks. Risk waits for a person. A failed step takes a backup path. The items are made up.
Identity. Each step runs as a named role. A tool outside that role waits.
Illustrative. Made-up data. No real clients, figures, or personal data. The model is one step among sources, checks, and a person’s approval.
The work is hard to see.
The failures that keep a service business stuck, and the system we put in their place.
Problem
Operations live in people’s heads and chat threads.
What we build
An explicit path: roles, routing, handoffs, and a checklist that runs the same way each time.
Problem
No single source of truth. The owner cannot see status.
What we build
One view of status, freshness, and cost. The cost view includes review time, not only the model.
Problem
Repetitive, document-heavy client work fills the day.
What we build
Intake and answers from your own sources, with a person when it should escalate.
Problem
Earlier automation was confident, and ungrounded.
What we build
Sources, citations, checks, monitoring, a human gate on risk, and a fallback. The model is the smallest part.
The layer around the model.
Sources, checks, limits, monitoring, and a person where the risk is high.
Business operating systems
How the business runs, written down: roles, routing, handoffs, checklists.
Workflow engineering
Routing, permissions, and the human gate stay in this layer. A tool may forget the previous call. This layer keeps the task, the role, and the approval.
Knowledge systems
Answers from your sources, with the source named. The system stops when an answer is not grounded.
Client automation
Intake, classification, answers, and follow-ups, with a path to a person.
Analytics and control
One screen for status, alerts, cost, and what needs a decision.
Five lines of work.
Diagnostic, a build sprint, or an operating retainer. How an engagement runs is the next section.
The system around the model
The layer around a model for one real workflow: routing, retrieval, checks, limits, evaluation, fallback, and a person approving risk.
Knowledge systems
Cited, source-backed answers from your own documents, built to stop when an answer is not grounded. Ingestion, retrieval, a citation on every answer, freshness, and a safe stop. One source set can feed more than one workflow. There is no per-seat meter.
AI product build, concept to production
A scoped product or feature, from the use case and the success criteria to a deployed asset. Model choice, the same reliability layer, and monitoring. Unit economics and data privacy are part of the build.
AI security and guardrails audit
An adversarial look at an existing model system, then guardrails, permissions, and human gates. We test for injection, jailbreak, extraction, leakage, and tools that reach too far. EU AI Act, NIST AI RMF, and ISO 42001 are the control checklist. The audit includes which role the agent is allowed to be. On the Commission page, high-risk duties are dated 2 December 2027. Transparency rules are described there as taking effect in August 2026. This list is not a certificate.
Business operating systems
Roles, routing, handoffs, dashboards, and document workflows, made into one system the owner can see. Controls come before automation. The other four lines plug into this. These roles are the identity the other four lines run as.
The layer sits on the tools you already run. It does not require one vendor's agent runtime.
Small steps, checked first.
Risk moves one step at a time.
A failed step does not count as done. It runs again, and a side effect must be safe to repeat.
About 1–2 weeks, fixed
Diagnostic
Map the systems, the code, and the operation. Find the bottlenecks. Return a prioritized plan, an evaluation set, and the list of which role may call which tool.
Scoped
Build sprint
From a prototype to a production-ready increment: designed, built, guarded, evaluated, and shipped. Shipped means an evaluation report on your cases and a spend cap, not a demo.
Monthly
Operating partner
A retainer. We run, monitor, and evolve the systems, with a clear service level. The retainer runs and watches the system. The runbook, the gates, and the evaluation set stay with you if the retainer stops.
What you leave with.
Something you can run, inspect, and operate.
The deployed system
Wired to your sources, tools, and workflow, behind guardrails and human gates.
System-design document
How it is wired: sources, retrieval, orchestration, permission boundaries, and where a person approves.
Evaluation report
Accuracy, refusal, and failure, measured on your real cases, with a trace you can debug. The report shows refusal and a failed step, not only a score.
Guardrails and human gates
The risk boundary in writing: what the system may do alone, and what waits for a person.
Monitoring and cost
A long-running session can carry a hard spend cap. When the cap is reached, the session pauses for a person. It does not keep spending. The cost view includes review time.
Fallback and runbook
What happens when a model or an API degrades, plus an operating runbook and support.
A person holds the risk.
Read-only first. Measured before it acts. A change in production waits for a person. Each run has a named role. A tool outside that role waits for a person.
Read-only first
A high-risk flow starts as review, synthesis, validation, or reporting, before any write.
Human gates
Money, secrets, infrastructure, and legal decisions stay behind an approval. The system proposes. A person decides.
Guardrails and permissions
A high-risk prompt can be held for an allow-or-deny decision before the model runs. Every denial is recorded.
Measured behavior
Tested on real examples. Benchmarks and failure cases are visible before the system is scaled.
Not tied to one model
Access to a model can be withdrawn overnight, by an outage, a safety review, or an export-control order. The layer keeps a documented fallback, so the workflow does not stop with one provider.
Governance as a checklist
EU AI Act, NIST AI RMF, and ISO 42001, used as controls to implement and to verify. The same checklist is a separate audit. On the Commission page, high-risk duties are dated 2 December 2027. Transparency rules are described there as taking effect in August 2026. This list is not a certificate. ISO/IEC 42001:2023: certification is voluntary. ISO does not certify organizations.
We are our own first client.
We run our own business on what we build.
Cited compliance assistant
Answers from controlled source material, with citations, and refuses when an answer is not grounded. If the official text is not in the source, the assistant stops and does not paraphrase.
Daily regulatory monitoring
Tracks regulatory changes across official sources every day, plus read-only market-data and analytics tooling. The morning note retries a failed fetch. It does not invent the text it failed to read.
Role routing and dashboards
Specialist work with handoffs, document workflows, and dashboards an owner can see, behind a public website. A handoff produces a decision card. The system fills it. A person confirms it.
A weekly engineering artifact
One measured artifact a week: an evaluation, a gate change, or a runbook page.
Sanitized only: no private data, credentials, client details, financials, or PII; no invented clients, metrics, or logos.
Before a build.
What does evaluation look like in production?
Return a prioritized plan, an evaluation set, and the list of which role may call which tool.
Which role may call a tool?
You get a map of which role may call which tool.
What stays if the retainer stops?
The runbook, the gates, and the evaluation set stay with you if the retainer stops.
Bring the operating problem.
Five lines are enough: the workflow you want to be reliable, your tools and where the sources live, what success looks like, the timeline, and any compliance or risk constraint. We reply with a clear next step, usually within 1–2 business days, typically a fixed-scope diagnostic.
Write to info@ajax.engineering