AI integration · built and shipped, not slide-decked
Ten years building production software, most of it at Datorama and then Salesforce, on systems that were not allowed to go down. I take AI out of the demo and put it inside the systems your business depends on, with the permissions, logging and cost controls that let you actually leave it running.
Why most of it stalls
Getting a model to give a good answer is the easy part now. The hard part is everything around it: which user is allowed to see which data, what happens when the model is wrong, what it costs at a thousand requests a day, and how anyone proves it is still working three months later.
That gap is engineering, not prompting. It is the same work I did for a decade on systems that could not go down, and it is why the projects I take on end up in production instead of in a folder called pilot.
What I build
Agents
Software that reads, decides and acts on its own. It pulls from wherever your information lives, works out what matters, writes back into your systems and tells the right person when something needs a human. Built around your process, not a template.
Automation
The recurring work that eats your team's week: intake, sorting, drafting, chasing, reporting. I map the process end to end, automate the parts a machine genuinely does better, and leave the judgement calls with the people who should be making them.
Your data
Your records, documents and history become something anyone can ask questions of in plain language. Scoped per user so people only ever see what they are allowed to see, and answered with a source attached rather than a confident guess.
MCP
An MCP server that lets customers and staff reach your business from the AI they already use. They ask their own assistant, it queries your systems with their permissions, and it answers with your data and your rules travelling alongside it.
How it works
I look at how the business actually runs and find the two or three places where AI moves a number you care about. You get a written plan with costs and risks, and it is yours whether or not we keep going.
One working thing, on your real data, in front of your real users. Deliberately small enough to throw away. This is where we find out if the idea survives contact with reality, before anyone has spent real money.
The prototype becomes production software: authentication, permissions, audit trail, spend limits, tests, monitoring. This is the step most AI projects skip, and skipping it is the reason most of them quietly stop.
Every answer traced, every shekel of model spend tracked, and a fixed set of test questions that runs on every change. If something makes the system worse, you find out in a day rather than in a customer complaint.
Track record
AI stack
Claude API · MCP · LangChain · LangGraph · RAG · pgvector · Embeddings · Prompt engineering · Evals · Tracing · Claude Code · Spec-driven development
Platform stack
Java · Spring Boot · Python · Kafka · Spark · Kubernetes · Docker · Terraform · PostgreSQL · Redis · Elasticsearch · Next.js · Jenkins · Splunk
Training
DevOps Experts
Distributed system design, scalability patterns and architectural trade-offs, studied alongside a decade of shipping the real thing. It is why I can tell you early which parts of a build will break at ten times the traffic, while changing them is still cheap.
Start here
I will tell you honestly whether AI is the right answer. Sometimes it is a database index and a form, and you should hear that from someone with nothing to sell you on the alternative. First conversation costs nothing and usually takes twenty minutes.
