Agentic AI work for me is production LLM applications: retrieval (RAG), tools, evals, and multi-tenant agent configuration. Not a chatbot demo. Models from OpenAI, Anthropic, and Google sit behind your data, your CRM, and your policies.
The model is the cheap part. The expensive part is grounding, tools, tenancy, and knowing when the agent is wrong. I build the application around the model so it cites your docs, calls your CRM, and fails closed on refunds, medical advice, or off-policy quotes.
On Use Cortana that meant a drag-and-drop studio: per-business knowledge bases, custom tools, first messages, and channel config. On Brisbane Gateway it meant booking agents that remember stay details and hand off to a real checkout link instead of hallucinating availability.
Evals are not optional. If I cannot grade a transcript, I am guessing. I start with one workflow and one knowledge source, then expand tools and tenants after the first path is boringly reliable.
Who it is for
- SaaS teams that need per-tenant knowledge and tools
- Ops teams drowning in docs the model has never seen
- Products that must call CRM, calendar, or internal APIs from a conversation
What you get
Grounded answers
Chunking, embeddings, and retrieval with memory so the agent cites your docs instead of inventing policy.
Tools the model is allowed to use
Function calling into GHL, booking, search, hang-up, with a human-readable audit of what ran.
Operator control
Prompts, first messages, and channel config in a studio. The same pattern as Use Cortana's agent builder.
How I work
Define failure
What must never happen (refunds, medical advice, off-policy quotes) before I pick a model.
Thin agent
One workflow, one knowledge source, evals on real transcripts.
Expand
More tools and tenants only after the first path is boringly reliable.
What this looked like for Brisbane Gateway Resort
Guests can book by text or voice. The agents collect stay details, show caravan, lodge, and cabin options (with images in chat), then pass a Newbook checkout link for payment. RAG memory keeps the conversation on the park, not on a generic LLM tourism spiel.
Proof
- Use Cortana: multi-tenant agent studio, per-business RAG, Hume and Twilio
- Brisbane Gateway: dual text and voice booking with RAG memory
- RAG write-up: vector memory patterns used in production agents
Case studies
- Use Cortana agent studio
- Vibe coding on Use Cortana
- Vibe coding Loom demo
- Use Cortana walkthrough Part 1
- Use Cortana walkthrough Part 2
- RAG in production
- Brisbane Gateway booking agents
Stack and docs
Related
Common questions
Is this ChatGPT wrappers?
No. The hard part is retrieval, tools, tenancy, and evals. The model is a component. I will say no if a form and a search box would be more honest.
Which model do you use?
Whatever fits latency, cost, and tool-calling quality: OpenAI, Anthropic, Gemini, or a mix. Routing is a product decision, not a brand loyalty.
Can non-engineers change the agent later?
That is usually the point. Studios, Retell configs, or GHL workflows so operators edit copy and hours without a pull request for every tweak.
Do you train custom models?
Almost never as the first step. Grounding, tools, and evals on a hosted model get you further than a fine-tune. If a domain really needs one, that is a later, explicit decision.
How do you stop the agent inventing policy?
Retrieval over your docs, tools instead of free-text side effects, and a fail-closed list (refunds, medical advice, off-policy quotes). If I cannot grade a transcript, I do not call it production.
Will my documents leave my account?
Embeddings and completions go to the providers you approve (OpenAI, Anthropic, Google, or a self-hosted option). I keep tenancy so one customer cannot retrieve another customer's knowledge base.
Can this sit on a product I already have?
Yes. The usual path is an API the model can call, a knowledge source, and a thin UI or channel. I do not require a rewrite of your app.
What does an eval actually look like?
A set of real transcripts or questions with pass/fail checks: cited the right doc, did not offer a refund, called the right tool. I run those before adding more tools or tenants.
How do you start an agentic build?
One workflow, one knowledge source, one channel. Expand after that path is boring. A discovery call is enough to see if this service or consultation is the better first step.