NoBroker co-founder Akhil Gupta, who can be the chief tech & product officer, and his small workforce of engineers have been attempting to enhance the productiveness of their buyer brokers for years. The drawback they confronted was easy: NoBroker’s brokers have to have interaction with numerous potential prospects relating to high-value buy gadgets – primarily homes and flats.
For the brokers to be efficient, they wanted to strictly comply with customary working procedures and have the appropriate data simply accessible to them. But contemplating the excessive attrition charges amongst customer support representatives, sustaining a top quality of service was a nightmare for everybody concerned, says Akhil.
The resolution to NoBroker’s conundrum was to depend on AI. That’s after they hit their second roadblock: AI options out there have been costly. So they determined to construct AI options themselves.
Their journey over the past half a decade have led them to develop a number of small language models (SLMs) and conversational analytic models for not simply their inner use however whilst standalone software program functions for a number of companies. “Some of these are task specific small language models and others special adaptations of open LLMs. These models have also become the foundations of our flagship conversational analytics tools called ConvoZen.AI. We feel that this is another startup which has taken birth within NoBroker,” says Akhil.
The main motive why their SLM-based AI models have been common is as a result of they will comprehend a number of Indian languages, even common lingua francas like Hinglish. They are additionally remarkably helpful. For instance, the AI models can summarise and spotlight errors made throughout an agent’s dialog with a buyer, give correct sentiment evaluation, and even present the agent with all of the contextual data they may have to serve their buyer higher. LLMs are costly Akhil stresses that he and his workforce by no means got down to create a language mannequin to rival the others accessible out there and that as an alternative they have been merely attempting to unravel an issue. “Lots of companies are trying to create their own base model. I don’t want to do it. I’m trying to solve a problem.”
And they have been ready to try this with just a few hundred reasonably powerful GPUs and a workforce of round 25 engineers.
“Our speech and language series of models are internally code named as Maya Series of Speech & Language Models. MayaConformer (a 50 million parameter model) is our state-of-the-art conversational transcription model which is based on conformer style architecture trained on 35,000+ hours of open data sets and 1,500+ hours of inhouse annotated data sets across multiple Indian languages. We have also used datasets available from Bhashini and CommonVoice.”
Akhil says that with the appropriate coaching information, small language models may be extremely efficient for slim, particular use cases. “You can build a model for a particular use case. You can also borrow it from large models that are out there. This is called Retrieval Augmented Generation modelling. For example, in the retail sector our word error rate may be 20%. But then with just ten hours of recording (of specific retail data), we can bring the word error rate down to single digits, only a small bit of curation is needed.”
Arabic comprehension When they began out, Akhil says they used to construct separate models for all languages, like Hindi, English, Tamil and Telugu, however now they’ve graduated to a single mannequin which might comprehend a number of languages. Akhil says that they’re assured that they will construct any mannequin for any language on the planet now -as lengthy as it’s for particular use cases. Which is why they even mastered Arabic.
“I was in Dubai doing a demo for Sheikh. We took a call, did a transcript (with NoBroker’s model) and I showed it to him, and I told him that I don’t understand Arabic and asked if he could please read it. His response was that it was perfect.”
For the brokers to be efficient, they wanted to strictly comply with customary working procedures and have the appropriate data simply accessible to them. But contemplating the excessive attrition charges amongst customer support representatives, sustaining a top quality of service was a nightmare for everybody concerned, says Akhil.
The resolution to NoBroker’s conundrum was to depend on AI. That’s after they hit their second roadblock: AI options out there have been costly. So they determined to construct AI options themselves.
Their journey over the past half a decade have led them to develop a number of small language models (SLMs) and conversational analytic models for not simply their inner use however whilst standalone software program functions for a number of companies. “Some of these are task specific small language models and others special adaptations of open LLMs. These models have also become the foundations of our flagship conversational analytics tools called ConvoZen.AI. We feel that this is another startup which has taken birth within NoBroker,” says Akhil.
The main motive why their SLM-based AI models have been common is as a result of they will comprehend a number of Indian languages, even common lingua francas like Hinglish. They are additionally remarkably helpful. For instance, the AI models can summarise and spotlight errors made throughout an agent’s dialog with a buyer, give correct sentiment evaluation, and even present the agent with all of the contextual data they may have to serve their buyer higher. LLMs are costly Akhil stresses that he and his workforce by no means got down to create a language mannequin to rival the others accessible out there and that as an alternative they have been merely attempting to unravel an issue. “Lots of companies are trying to create their own base model. I don’t want to do it. I’m trying to solve a problem.”
And they have been ready to try this with just a few hundred reasonably powerful GPUs and a workforce of round 25 engineers.
“Our speech and language series of models are internally code named as Maya Series of Speech & Language Models. MayaConformer (a 50 million parameter model) is our state-of-the-art conversational transcription model which is based on conformer style architecture trained on 35,000+ hours of open data sets and 1,500+ hours of inhouse annotated data sets across multiple Indian languages. We have also used datasets available from Bhashini and CommonVoice.”
Akhil says that with the appropriate coaching information, small language models may be extremely efficient for slim, particular use cases. “You can build a model for a particular use case. You can also borrow it from large models that are out there. This is called Retrieval Augmented Generation modelling. For example, in the retail sector our word error rate may be 20%. But then with just ten hours of recording (of specific retail data), we can bring the word error rate down to single digits, only a small bit of curation is needed.”
Arabic comprehension When they began out, Akhil says they used to construct separate models for all languages, like Hindi, English, Tamil and Telugu, however now they’ve graduated to a single mannequin which might comprehend a number of languages. Akhil says that they’re assured that they will construct any mannequin for any language on the planet now -as lengthy as it’s for particular use cases. Which is why they even mastered Arabic.
“I was in Dubai doing a demo for Sheikh. We took a call, did a transcript (with NoBroker’s model) and I showed it to him, and I told him that I don’t understand Arabic and asked if he could please read it. His response was that it was perfect.”
