JDM System
The architecture behind every AI Employee.
Business in, outcome out. In between, one Core that turns models, data and automation into agents, knowledge and workflows you can open up and inspect.
The stack
Seven layers, from the chat window to the server.
Select a layer and the Core slides it out of the stack. Problems in AI systems rarely live in the model; they live between these layers.
Where customers and staff meet the system: WhatsApp, web, portals and review queues.
- Conversation design
- Portals
- Review queues
Radd Core
Every role runs on the same eight parts.
A Sales AI and a Support AI differ in configuration, not in code. Fix something once and every role gets the fix.
- IdentityRole, tone and boundaries
- KnowledgeYour facts, searchable
- MemoryConversation state
- ToolsBookings, leads, handoff
- ChannelsWhatsApp, web, more
- GuardrailsConfidence and escalation
- LoggingEvery decision recorded
- MeteringUsage and cost tracked
Knowledge layer
Your knowledge, usable by the system.
A model knows the world. It does not know your rates, your policies or what you promised a customer last week. The knowledge layer connects the information your people already use, keeps it current, and lets every answer be traced to its source.
AI = model + context + knowledge + tools + workflow + controls
- Documents
- Website
- CRM
- FAQs
- Policies
- Products
- Databases
- Emails
- Internal notes
Agent runtime
Nine steps between a message and an action.
A prompt answers. A system acts, checks itself and leaves a record. The last step is always available: hand the work to a person.
A message, an email, a form or an event from another system.
Observability
If it happened, it is in the trace.
A chat log tells you what was said. The trace tells you why: which source, which rule, which tool and which person. This is the record an owner reads to audit a conversation in a minute.
- Received WhatsApp messageFrench · booking enquiry
- Retrieved room availabilityCalendar · 12 to 14 March
- Applied rate ruleWeekend rate · breakfast included
- Generated replyTwo options · 46 words
- Booking request createdStatus: awaiting guest
- Transfer flagged for receptionRule: transfers confirmed by a person
Running
- Execution
- 6 steps
- Latency
- 2.3 s
- Tools used
- Calendar, booking
- Human intervention
- 1 handoff
- Outcome
- Completed
Illustrative trace for a fictional guesthouse. It shows what the system records; it is not a client measurement.
Evaluation
Every failure becomes a test case.
Every change is re-evaluated before it reaches a customer, in all four languages. We publish scores only when they come from real deployments.
- Task success
- Did the work get done as specified?
- Response quality
- Correct, complete, in the right language and tone.
- Escalation rate
- How often a person was needed, and whether they should have been.
- Tool success
- Calls that ran and returned what was expected.
- Error rate
- Wrong facts, invented answers or rule breaks.
- Latency
- Time from request to useful answer.
- Cost
- Model and infrastructure cost per task.
- Customer outcome
- What happened next for the person on the other side.