Taking on new projects · Sunshine Coast, QLD
Azul Labs is named for azulejos: glazed tiles that are plain one at a time, and a picture once they’re laid.
We build AI the same way, from the Sunshine Coast. Agents, local and fine-tuned models, and evals, each piece small enough to check and yours to keep. We help startups and scale-ups figure out what’s actually worth building, and then build it. And when data can’t go to the cloud, we can run the work privately for you.
$ uvx clai -a support_agent:agent "Where’s order 4417?"
Order 4417 shipped on Tuesday and should arrive Friday.
$ uv run evals/support_agent.py
92.5% pass 1.8s p50 local model
Works with
Services
Most AI advice is hype. We help you figure out where AI actually makes sense for your business, build the integration or agent, and show your team how to use it. Claude, APIs, workflow automation.
Some businesses can't send their data off to a cloud AI provider. We set up open models on your own servers or cloud account and hook them into the tools you already use. We check the weights against what the publisher released, and run them sandboxed. See the local AI set-up.
When a good prompt isn't enough, we fine-tune open models on your data so they learn your domain, your formats and your tone. Evals before and after show whether the tuned model is actually better, and the tuned model is yours to keep.
Changed a prompt or swapped models and not sure if it helped? We write evals for agents so you can check. Test cases come from real conversations and run in CI like any other test.
FastAPI and Pydantic are our go-to stack. Clean, fast, and easy to maintain. We build backends that don't fall over when you actually start using them.
Apex, Lightning, CI/CD, infrastructure as code. If your Salesforce deployment pipeline needs work, we can help.
AI sovereignty
If your business runs on AI, the models, prompts, evals and data behind it are part of what makes you valuable. When all of that lives inside someone else's platform, you're renting it. We set things up so the important parts belong to you.
Code, prompts, datasets, evals and any fine-tuned models are handed over to you, not left in our accounts or locked in a vendor's.
Open models and standard tools mean you can change providers, or bring everything in-house, without starting again.
On your own servers, in your cloud account, or on Claude where it's the right tool for the job.
Private work
Some jobs shouldn't go anywhere near a cloud AI provider. If you need documents analysed, a model run over sensitive data, or a fine-tuned model tested, we can run it on infrastructure we run ourselves and hand back the results.
Ask about private workYour data isn't sent to OpenAI, Anthropic or any other AI provider.
Useful for analysis or a one-off job where buying hardware makes no sense.
We agree what runs, and what happens to your data afterwards, before anything starts.
If something needs to keep running, we'll usually set it up on your own hardware instead.
How we work
You tell us what you're trying to fix. If AI isn't the right tool for it, we'll say so.
You get something to try early on, then we build it out properly. Claude, a local model or both, depending on cost and how sensitive the data is.
With evals you can run yourself and code your team can read and maintain.
About
Azul Labs is an AI engineering practice on the Sunshine Coast. We build agents, set up and fine-tune local models for teams that need to keep data in-house, and write the evals that show an agent works before it goes live.
We work best with startups and scale-ups that have a real problem to solve: building something that works, and explaining it in plain English.
We also maintain Forcebench, an open benchmark of AI models on Salesforce work.
Leo Alves
Founder. 15 years building software across AWS, Salesforce and Python, and now AI with Claude and local models. LinkedIn · Working notes
FAQ
Yes. Open models are now good enough for a lot of business tasks. We can set one up on your own servers with vLLM or MLX, so your data stays on your network.
Yes. If you need documents analysed or a model run over sensitive data, we can do it on infrastructure we run ourselves, so nothing is sent to OpenAI, Anthropic or any other AI company. For anything long term, we'd usually set it up on your own servers instead.
Partly. We check the model files against the hashes the publisher released, and use formats that can't carry hidden code, so you know you're running the weights that were actually published. Nobody can prove a model has no hidden behaviour trained into it. So we also run models and agents in a sandbox, with no route to the internet or to files they don't need, and use evals to check how they behave.
Evals are tests for an AI agent. They run it against real examples and score the answers, so you can tell whether a prompt change or a new model made things better or worse. If an agent talks to customers or makes decisions, you want them. We write ours with Pydantic Evals.
When a good prompt and a few examples still don't get a model where it needs to be, or when you want a small model that does one job really well. We run evals on the base model and the tuned one, so you can see whether the tuning paid off.
You do. Code, prompts, datasets, evals and any fine-tuned models are handed over to you. Open models come with their own licences, so we check those fit how you plan to use them.
No. About 25% of our work is handled by Claude and the other 75% by local models. Not every percent is equal, though. That 25% goes to the tasks local models can't handle yet.
Yes. Apex, Lightning, CI/CD and deployment pipelines on the Salesforce side, and FastAPI and Pydantic backends on AWS for Python.
The Sunshine Coast in Queensland, Australia. Azul Labs Pty Ltd is the business name.
Got a project in mind, or just want to know if running AI locally makes sense for you? Happy to chat.
leo@azl.au