We teach this publicly
Full-day Embabel and Spring AI workshops at W-JAX and Workshop-Tage, plus sessions at Spring I/O, JAX, and JUG Switzerland.
Agentic AI Engineering
A convincing demo is not yet an operable system. CTOs and engineering managers need evidence on quality, clear tool permissions, bounded cost, failure handling, and an operating model. We build JVM agents with Spring AI and Embabel, combining goal-oriented planning with evaluations, observability, and explicit safeguards.
Fixed-scope discovery and first production agent. Final price depends on scope and integration surface.
Anonymised outcome: a Swiss enterprise LLM proof of concept became an agent using goal-oriented planning, with evaluation harnesses, guardrails, and production observability.
Not sure you are ready? Take our free AI readiness checklist – 35 questions, no sign-up, printable.
The problem we are usually called about
Typical concerns include evaluation evidence, cost ownership, integration, data protection, and the operating model.
-
The pilot is not ready to operate
The proof of concept handles its demonstration path, but the team lacks representative evaluations, reproducible traces, and a way to investigate failures.
-
Cost ownership is unclear
Token spend is unpredictable and invisible. There is no budget per request, no model routing strategy, and no way to tell an expensive call from a wasteful one.
-
It sits outside the estate
The prototype uses a separate runtime from the production estate. That may be justified, but it also brings another deployment path, operational responsibility, and security review that the architecture must account for.
-
Data protection was addressed too late
Data residency, auditability, and retention were not designed in. Addressing them after providers, data flows, and tool boundaries are fixed can require substantial rework.
What we build
Agentic systems on the JVM, using Spring AI for model access and Embabel for goal-oriented planning – the agent framework created by Rod Johnson, founder of the Spring Framework.
-
Agents with goal-oriented planning
Embabel can plan dynamically over typed actions and explicit conditions. The selected path can be traced and reviewed; LLM-backed actions remain non-deterministic and require evaluation, validation, and operational controls.
-
Grounded retrieval
RAG pipelines over your own content: chunking and embedding strategy, vector-store selection, retrieval evaluation, and source references that reviewers can inspect.
-
Tool and system integration
Function calling and Model Context Protocol clients and servers that let agents act on your real systems, with the same authorisation boundaries your APIs already enforce.
-
Evaluation and guardrails
Regression suites for non-deterministic behaviour, golden datasets, output validation, and human-in-the-loop checkpoints on the steps that carry real business risk.
-
Observability
Micrometer, Actuator, and OpenTelemetry instrumentation so latency, token spend, failure modes, and drift show up on the dashboards your operations team already watches.
-
Cost and model strategy
Routing between local and hosted models, caching, prompt-size discipline, and per-request budgets – so unit economics work at production volume, not just in a pilot.
How an engagement runs
Fixed-scope, low-commitment start. You can stop after any stage with something useful in hand.
-
Assess
Two to three days. We review your use case, data, estate, and constraints, and assess whether an agent is the right approach.
-
Design
Goals, typed actions, tool boundaries, data flows, model strategy, and data-protection and governance requirements – written down and reviewed with your architects before code exists.
-
Build
We implement the first agent end to end, pairing with your engineers throughout so the knowledge lands in your team rather than leaving with us.
-
Harden and hand over
Evaluations, guardrails, dashboards, runbooks, and a working session with the people who will carry operational responsibility. We then reduce our involvement.
What you get
A running agent, evaluation evidence, and operating material your team can use.
- A running agent in your environment, integrated with your systems
- An architecture document covering goals, actions, tools, and data flows
- An evaluation harness with golden datasets and regression suites
- Guardrails and human-in-the-loop checkpoints on high-risk steps
- Observability dashboards for latency, cost, and failure modes
- Documented data-protection and model-routing decisions for your privacy and compliance specialists
- Runbooks and a handover session with your operations team
- Working examples and handover sessions that help your engineers extend the agent
Who this is for
-
CTOs and engineering managers with a JVM estate
You run Java and Spring in production, the board wants AI in the roadmap, and you need it done without an unjustified parallel stack, a new on-call rota, or unexpected approval work.
-
Teams with a stalled pilot
Something works in a notebook or a demo branch, but the team still needs representative evaluations, cost controls, security review, and an operating model.
-
Architects who have to sign it off
You need an agent design you can defend in a review board: an inspectable plan, evaluation evidence, clear failure modes, bounded cost, and documented data flows.
Why us
Our work combines production Java experience with public teaching, technical writing, and hands-on work with Spring AI and Embabel.
We publish what we learn
Patrick writes for Java Magazin about Embabel, Spring AI, and MCP servers, with examples readers can inspect and challenge.
We build the tooling
We maintain an open-source Spring Boot starter that visualises Embabel agent workflows live through an Actuator endpoint.
We understand the surrounding system
Java Champion, Oracle ACE, and VMware Certified Spring Instructor. More than 20 years in enterprise Java helps us integrate AI work with the architecture, security, delivery, and operations your team already owns.
More on the people behind this on our company page, or take the Spring AI & Embabel training if you would rather build it yourself.
Frequently Asked Questions
Why Java instead of Python?
Python can be a sound choice. When the relevant domain logic, authentication, data access, and operational controls already live in a Java and Spring estate, keeping the agent there can avoid a second runtime and duplicated integration work.
What is Embabel?
An open-source agent framework for the JVM created by Rod Johnson, founder of the Spring Framework. It can plan dynamically over typed actions and explicit conditions, making the selected path inspectable. LLM-backed actions still require evaluation and controls.
We already have a proof of concept.
We assess what you have, retain what is useful, and identify the work needed for evaluation, cost control, latency, error handling, security, and auditability.
Which model providers?
OpenAI, Azure OpenAI, Anthropic, Google, Mistral, and local models via Ollama. Spring AI's portable abstractions keep provider choice open – which matters for Swiss data residency and for cost control as models change.
How do you support data-protection requirements?
We document proposed data flows, provider and region choices, logging, and retention in the first architecture session. Your data-protection and legal specialists confirm the applicable requirements. Local models remain an option where the approved design requires data to stay in your estate.
What does it cost?
An engagement starts with a fixed-scope discovery and first production agent. The final figure depends on scope, integration surface, and how much your own team takes on. Book a call and we will scope it with you.
Plan the path from pilot to production
Book 30 minutes with Patrick. Bring your use case and your constraints. We will assess whether it is worth building and what an appropriate first stage would cover.