Hello Chitti

This is a Freshsales Marketplace app built during Appathon.

Goal: To build an voice assist system for agent.

Current Implementation: We have build an agent facing widget via marketplace, which can be installed on any of the freshworks product which have integration with marketplace. Widget is able to perform both on text as well as voice input.

WorkFlow:

1. Widget will collect the agent query, forward it to System42 via an API call, which will further be routed to ML to tokenize it and extract entities and send the response back to system42.
2. System42 here is a decision making system, which will do phonetic match, cleaning the response, adding the default operator, verify the confidence, etc.
3. Post that System42 will send the response back to widget. Marketplace app will forward the request to specific route and apply the filters as well.

ML role.

We have used RASA as infrastructure level to build our ML models. We have used only the NLU part for intent detection.

Entities for which we trained on

1. type (model_name / table name on product)
2. Operand (Attribute_name / field name on product)
3. Operator (operation that need to be perfomed)
4. values (Actual data) (could be name/date/number/currency)

We also have smalltalk intent

SYSTEM42 role.

We have used system42 as the decision making system. We are exposing an api to the marketplace app.Through this we will collect the customer query and proxy it to ml to identify the entities. Once the entities are collected, System42 will take certain decisions.

1. In the query the type should always be there. If not it will give unknown intent.
2. If Operand is there, then value should also be there.
3. If operator is not there, but operand and values are there, then it will insert the default value of operator if we have, otherwise mark it as unknown intent.
4. It will also take decision on confidence. System42 can tune intent confidence to get more better result.

Built With

• css
• html
• javascript