AI agents & assistants
Software that answers, decides and acts inside your business.
Every engagement fits one of these four shapes. Open the one closest to your situation.
We build AI agents that sit on top of your real data — catalogue, CRM, tickets, documents — and handle the conversations and decisions your team repeats a hundred times a day. Not a chatbot bolted onto a website: an agent with tools, permissions and an audit trail.
- 01Your team answers the same forty questions every day across WhatsApp, email and social inboxes.
- 02Answers differ depending on who replies, and nothing is logged in one place.
- 03Generic chatbots hallucinate because they were never connected to your actual data.
- 04Nobody can tell you what the assistant did last week or why it said what it said.
What we build
Software that answers, decides and acts inside your business.
- OpenAI / Anthropic / Gemini models
- Vector search over your own corpus
- Function calling with typed schemas
- WhatsApp Business Cloud API
- Postgres + row-level security
- Queue-based async workers
Grounded retrieval
The agent answers from your documents, catalogue and records — with citations back to the source row or file, not invented text.
Tool use and actions
Check stock, create an order, book a slot, escalate to a human, open a ticket. The agent calls real functions with real permission boundaries.
Channel coverage
WhatsApp Business API, web widget, Messenger, Instagram, email. One brain, several front doors, one conversation history.
Human handover
Confidence thresholds and explicit triggers route the conversation to a person with the full context attached.
Evaluation and guardrails
A regression suite of real questions runs on every prompt change, so quality is measured rather than hoped for.
Observability
Every message, tool call, cost and latency is logged and searchable. You can replay any conversation.
Questions we get
Will it invent answers?
It answers from a retrieval layer built on your own content, and refuses or escalates when confidence is low. We test that behaviour explicitly before launch.
Which model do you use?
Whichever fits the task and budget. The model sits behind an interface, so we can switch providers without rewriting the system.
Who owns the data?
You do. It lives in your database and your storage, and we can deploy so that no conversation content is retained by the model provider.
Describe the problem. We'll tell you what should actually be built.
Sometimes the answer is a full platform. Sometimes it is one integration and a rule engine. We will say which.
No specification needed — describe the problem in a few lines.