Skip to content
Huematte

Pre-trained on how a domain talks, before it meets your customers.

Every Huematte agent starts with an intent map of its vertical: the questions customers raise, the many ways they phrase them, the records each one needs, and the edge cases that trip up general assistants.

Intent map, sample entries

3 domains
Customer phrasingUnderstood asRecords it needs
“wrong fit, need bigger” ExchangeOrder, delivery date, stock
“parcel says delivered but nothing came” Missing deliveryCourier scan, address
“why is this month higher” Charge explanationInvoice lines, plan history
“card nahi chal raha” Failed paymentPayment attempts, grace period
“can't make it thurs” RescheduleBooking, clinician slots
“is the doctor in on sunday” AvailabilityClinic hours, rota

What it means

Four layers the agent already knows on day one.

Pre-training is not a longer prompt. It's a structured understanding of the domain that your policies then plug into.

01

Intents

The recurring requests of the vertical, such as exchanges, proration questions or reschedules, each with a clear definition of done.

02

Phrasing

Short, vague, misspelt and mixed-language messages, including Hinglish, mapped to the right intent without a clarifying round trip.

03

Required context

Which records settle each intent: the delivery date for a return, the billing date for proration, the clinician rota for a slot.

04

Edge cases

Partial returns, annual plans changed mid-term, same-day reschedules, and other cases where the obvious answer is wrong.

How it's validated

Measured on held-back examples, then on your own history.

Training is only useful if it holds up on conversations the agent hasn't seen. We check it twice: once per domain, and again on your data before launch.

01

Held-back domain sets

Each domain agent is scored on example conversations kept out of training, covering every intent and the known edge cases.

02

Intent accuracy

Did the agent understand the request correctly? Misreads are reviewed by intent so weak spots are fixed at the source.

03

Answer review

Is the answer correct under the policy it cites? Correct intent with a wrong rule still counts as a failure.

04

Handoff correctness

Did the agent hand off every case it should have, and only those? Both missed and unneeded handoffs are tracked.

What stays separate

Domain knowledge is shared. Your data is not.

The patterns of a domain are learned from curated domain examples. What belongs to your business stays in your workspace.

01

Your conversations

Used to test and improve your own agent only, never to train an agent for another customer.

02

Your policies

Loaded into your workspace for grounding. They shape your agent's answers and no one else's.

03

Your records

Read at the moment of a conversation through the access you grant, not copied into training data.

04

Your corrections

Fixes your reviewers make apply to your agent immediately and stay with your workspace.

Questions

Domain Training

Next: Policy & Knowledge Grounding
Do we need to provide training data?

Not for the three pre-trained agents. You provide policies and record access for grounding, and past conversations for the test stage. Custom domain agents do need labelled examples.

Which languages are covered?

English, plus the everyday mixed English used by customers in India, such as Hinglish. Tell us if your customers write in other languages and we'll confirm coverage before you start.

What if our business has intents the map doesn't include?

They can be added to your workspace during setup. Until then, the agent hands those conversations off rather than guessing.