- Contact centers face significant inefficiencies and high operator stress primarily due to "after-call work" (ACW), which can consume nearly as much time as the customer call itself.
- The WiseOps solution introduces a four-stage, low-latency AI pipeline to process messy audio streams, extracting structured and actionable business intelligence from customer interactions.
- By automating call summarization and data extraction, the system significantly reduces ACW time by 50%, improves data quality, and mitigates operator cognitive load and turnover.
Contact Center Voice AI: Low-Latency Intelligence Extraction from Messy Audio Streams — Dippu Singh
- After-call work (ACW) is a critical bottleneck in contact centers, often taking 6.3 minutes for an average 6.5-minute call, leading to high stress and operator turnover.
- The core engineering mission is to use AI to mechanize call summarization and data extraction from audio streams, aiming for a 50% or more reduction in post-processing time.
- The solution is a four-stage low-latency pipeline: Voice Capture, Speech-to-Text (STT) Engine, Generative AI Core, and Customer Data Sync.
- Voice Capture must perform real-time noise filtering, audio normalization, critical channel mapping (separating agent and customer audio), and early-stage Personally Identifiable Information (PII) masking.
- The STT engine requires >90% accuracy, leveraging advanced acoustic modeling, domain-specific dictionaries (e.g., "term life" vs. "term"), inverse text normalization, and auto-punctuation for effective LLM input.
- The Generative AI Core uses an orchestration layer with prompt templates and few-shot learning for structured outputs (e.g., bullet points for inquiry/action), a reasoning layer for intent classification, and a trust layer for token optimization and hallucination checks.
- A Customer Data Sync layer acts as an API gateway to map AI-generated JSON output to CRM fields, incorporating a human verification step before final data commit.
- Implementation resulted in a 50% reduction in ACW (from 6.3 to 3.1 minutes), significantly improved data quality and standardization, and reduced operator cognitive load.
ACW — After-Call Work; the post-call administrative tasks operators complete, such as note-taking and summarization.
STT — Speech-to-Text; an engine that transcribes spoken audio into written text.
Generative AI — Artificial intelligence models capable of producing new content, such as text summaries or images, from given prompts.
LLM — Large Language Model; a type of generative AI trained on vast amounts of text data to understand and generate human-like language.
PII — Personally Identifiable Information; data that can be used to identify a specific individual, like credit card numbers or passwords.
Hallucinate — In the context of AI, when a model generates information that is plausible-sounding but factually incorrect or fabricated.
Few-shot learning — A machine learning technique where a model can learn a new task from a very small number of examples.
Token optimization — Techniques used to reduce the number of tokens (basic units of text) processed by an LLM, often to manage costs or latency.
Schema mapper — A component that translates data from one structured format (schema) into another, ensuring compatibility between different systems.
Inverse text normalization — The process of converting written numerical or symbolic representations (e.g., "$5,000") back into their standardized numerical format from spoken forms (e.g., "five thousand dollars").
Hello everyone, welcome to the AI Engineer 2026 online track. Good morning, good afternoon, good evening, depending upon your time zone. My name is Deepu Singh and I lead the initiatives in the emerging data technologies and AI architecture at Fujitsu North America. And today I will be giving you or maybe provide your deep dive into a very specific but a very highly impactful engineering challenge which we have encountered in our contact center. So with that, let me just quickly share my screen and get this going. All right, so this is the topic of the discussion. WiseOps, a fine, low latency, intelligence extraction from messy audio streams. When we talk about, you know, generative AI, we often focus on clean text inputs but if you go into the real world specifically in the customer service or maybe contact centers specifically, the data does not start as a clean text, right. It starts as messy, quite overlapping, sometimes emotionally charged and it may have multi-channel audio streams also within that, right. And today we will explore the technical architecture which is required to capture that audio, process it with ultra-low latency and use the generative AI to extract structured, actionable business intelligence out of it. So with that, let's get started with this. So here is our roadmap for the next 25 minutes. First, we will stage with what the current challenges are. We have to understand the operational realities and the intense human bottlenecks in the modern contact centers to understand why this engineering matters. Second, we will walk through the solution which is provided. I will break down our high-level architecture into four key components detailing how we move from raw audio to a structured JSON using advanced summarization workflows. Thirdly, we will look at the key outcomes specifically like how these technical implementations they translate into hard ROI and operational impacts. And finally, we will discuss the roadmap ahead and being transparent about the current engineering constraints which we face now and then and where we are taking this technology and the roadmap ahead. So to engineer a great solution, we first need to deeply understand the problem. So let's look at the current state of the contact center operations. So contact centers, they are the front lines of the customer experiences but structurally they are breaking under the pressure. If you look at the industry data, over 50% of the contact centers, they identify hiring, training and productivity as their most critical barriers. Why? Because the job is incredibly difficult. When we analyze the reasons why the operators they leave for the other professions, high stress is the number one factor as part of it. Now operators are expected to handle complex customer emotions, navigate the multiple customer data platforms or the CRM systems involved into it and also document everything perfectly. So if you think about this, this leads to massive retention problem. We are caught in a negative spiral where understaffing leads to higher stress for the remaining operators which leads to height and over and that ultimately implies to more understaffing. So to break this cycle, we just cannot hire more people. We have to fundamentally engineer the stress out of the workflow. And the most clearing inefficiency in this workflow is something called after-call workflow or ACW, which we are going to discuss as part of this slide. According to our baseline studies, the average contact center called it typically last about like 6.5 minutes. However, the average post-processing time where the operator types of the nodes, some rises the call and selects the disposition codes, it takes like almost 6.3 minutes. So this is nearly like one is to one ratio we are talking about. And this also implies operators are spending almost as much time doing administrative data entry as they are actually talking to the customers. Furthermore, because the summarization relies on an individual operator's memory and the writing skills, the data quality is highly inconsistent. Our core engineering mission here was clear like use AI to target the after-call work which is ACW. If we can mechanize the summarization and the data extraction, theoretically we can reduce the post-processing time almost by 50% or even more. And this shifts the enterprise focus from merely just handling the calls to actually analyzing the voice of the customers for the business growth which we are looking out every now and then. So how do we engineer that shift? Let's deep dive into the technical solution and the architecture we built to solve this. So we designed a four stage low latency pipeline to transform the conversational audio into a structured business intelligence with minimal human intervention. And it starts with voice capture which is tapping into the telephony system to extract raw high fidelity audio streams. And that flows into our speech to text engine STT which is responsible for high accuracy transcription. Next is the brain of the system, the generative AI core. This is where we do the heavy lifting of the intent recognition and summarization. And finally, the customer data sync layer which translates those AI insights into API calls to update the customer data or CRM data automatically. Let's look at the engineering under the hood for each of these components in detail. So the first component is the voice capture in AI. The typical rule is garbage in equals garbage out. So if your audio intake is flawed, the LLM will hallucinate later on. We do real time audio intake. So applying noise filters to strip out the back office chatter or anything of those sort which is creating an attenuation is important and normalize the audio level is very, very important for us. Crucially we perform these channel mapping. We absolutely must split the stereo audio to isolate the agent on one channel on the very left and the customer on the other side which is the very right. So if you mix them into a single monotrack, like kind of overlapping with each other, the AI will struggle to figure out who said what and thereby reigning the entire downstream summary. So it's important that we have that channel mapping intact and it separates which who is what and who is saying what right. Finally, we apply a security layer because sometimes the audio streams can contain the credit card numbers or maybe passwords or anything which is personally identifiable information PII. So we utilize buffer management and early stage PII masking technique so that the sensitive data it never hits the LLM memory banks whenever we are moving ahead in the channel. Now next the audio it hits the speech to text engine for generative AI or the LLM to summarize the data effectively or your response effectively. We found that the speech to text the STT accuracy must be above 90%. We utilize advanced acoustic modeling to map the phone maze and filter out any regional dialects. We then apply the language logic utilizing domain specific dictionaries just for example if it's an insurance agent the speech to text engine STT needs to know the difference between a term life and a term right. Both of them are very close to each other but there should be a difference between them. Finally post processing is also vital. We use inverse text normalization and auto-punctuation. For example if a customer says $5,000 the speech to text engine it should must output that into numerical fashion and this numerical formatting it drastically improves the LLM's ability to extract the entities in later point in time. Now we reach the generative AI core we are not just throwing a raw transcript at LLM but we are also asking it to summarize it like we use a highly orchestrated approach in this case. In our orchestration layer we use specific prompt templates our setup showed that if you just ask an LLM to some riser call it output some messy narrative paragraph. So instead we use few short libraries to instruct the LLM to output separate bullet points. One list for customer inquiry and a separate list for operator's action. Then comes the reasoning layer. In the reasoning layer we extract the intent. We provide the LLM with a predefined list of customer called reasons like cancellation or new application or any kind of claim status and instructed to classify the transcript and output the reason why it choose that specific classification. And finally the trust layer where we apply the token optimization to keep the latency low and runs the automated hallucination checks to ensure the generated summary is strict and it is grounded in the transcript. Now the final technical hurdle is getting this beautiful data back into hands of the business. Our API gateway acts as a schema mapper. It takes the JSON output from the LLM and maps the field like customer intent or resolution status and directly to some corresponding fields based on the company's CRM system or any customer data which we have laid out via some REST APIs. So we don't remove the human entirely in this. We use a verification step in between the operator. It sees the AI generated summary auto-populated on the skin. They do a quick visual field validation. Make some minor edits if necessary and then just click the confirm. Samultaneously this structured data it flows into our business intelligence models aggregating the voice of the customer data for management dashboards and automatically you know flagging the candidates for new FAQ data entries. Now to tie the architecture together this is the linear workflow logic of the data pipeline. We take the raw transcript complete with the time indexing confidence scoring and denoising. We pass it through the speaker separation because we split the stereo channels in the step one. We can easily stitch the dialog together logically like customer said x agent said y. Then we move to the context deduction where the lm spots the entities like account numbers or product name or customer name run the sentiment analysis and recognizes the intent. The final state is the structured output so instead of a wall of text the system it outputs like clear and a clean JSON schema matching some predefined customer data or maybe CRM templates which we have in the enterprise and then categorize neatly into bullet points and this strict formatting is what turns an unstructured conversation into a database ready asset. So what happens when we deploy this architecture in a real contact center environment. Let's look at the outcomes. The operational impacts out of the implementation were quite immediate and highly measurable. Look at the ECW time. Under the manual operation after average after call work was 6.3 minutes powered by our AI workflow that dropped to like 3.1 minutes which is like almost 50% reduction in the processing time. If you calculate that across 500 seats handling thousands of calls a day you are looking at a massive operational saving the equivalent of almost reclaiming dozens of full time headcounts purely from efficiency standpoint. The next aspect is of data entry quality it moved from a highly subjective and a variable to highly standardized and uniform output. The inquiry categorization or the call reason tagging it moved from being dependent on an operator mode or memory to being strictly logic based resulting in a highly consistent voice of the customer data set for the management. And ultimately by removing these repetitive administrative burden of typing out notes we reduce the cognitive load on the operators thereby stabilizing the operations and directly combating stress which is linked with the staff and this ultimately reduce the turnover we identified at the very beginning of our discussion. While the results were fantastic the engineering work is never done and let's talk about the constraints we face in our roadmap for the future. We are currently navigating three main constraints first is the source to text accuracy, STT accuracy. The entire generative AI summary it relies on the transcript so if the STT engine fails to pick up heavy accents or poor audio quality the LLM has nothing to work with so the STT optimization is a continuous battle for us. Second is the initial setup cost while the long term ROI is massive. The initial consumption of the API tokens especially running complex LLM reasonings on long 20 minute transcripts can be costly and especially during the initial scaling those are tough ones and we are constantly working on token optimization techniques to bring this number down. The third is the security and compliance handling PII any kind of sensitive information in audio streams is very complex and ensuring robust masking before the data hits Claude endpoint is a strict requirement and because of this we add some layers and ultimately reduces the latency and also adds some overhead from architectural standpoint so we are still figuring it out how we can reduce those extra layers and make it much more robust component-wise. So to address these constraints and push the boundaries of what's possible our roadmap is currently broken down into three phases. Phase 1 it focuses on explainable AI. We want to move beyond just summarizing calls to actually you know coaching the operators. We are engineering the systems to analyze the audio post call and provide operators with instant private feedback on their either soft skills or empathy level or any kind of accuracy from a information standpoint. The phase 2 it targets the predictive staffing by taking the massive amount of categorized intent data we are now capturing. We can feed the same exact data intent data into a time series analytics and this will allow the workforce management to accurately forecast the call volumes spikes based on those specific topics and thereby optimizing the shift scheduling. The phase 3 is perhaps the most important for human well-being combating the customer harassment. Contact center agents they mostly face an increasingly amount of verbal abuse and we are developing a low latency sentiment and acoustic analysis that can detect when a customer becomes abusive. Ultimately the system can you know trigger some triggers I mean analysts or something which is important from notification standpoint to supervisor or anyone who is there from the management upper management standpoint or maybe seamlessly transfer the call to an AI voice agent just to protect the human operator mental health in case of these tough conversations. All right so by applying these rigorous engineering techniques to the messy audio data we can definitely transform the contact centers from cause centers of like high stress into a highly efficient and intelligence gathering engines that protect their work forces that's the whole idea of having this right. Thank you so much for your time and listening to me. I have included my QR code which I will just flash in here that you can grab me over LinkedIn and please feel free to connect if you would like to discuss the architecture or any prompt engineering strategies related to the discussion which we had in here and happy to connect and make things work things out between us and we can have more conversations. So thank you so much for listening to me and have a good one. Bye.
TL;DR
- Contact centers face significant inefficiencies and high operator stress primarily due to "after-call work" (ACW), which can consume nearly as much time as the customer call itself.
- The WiseOps solution introduces a four-stage, low-latency AI pipeline to process messy audio streams, extracting structured and actionable business intelligence from customer interactions.
- By automating call summarization and data extraction, the system significantly reduces ACW time by 50%, improves data quality, and mitigates operator cognitive load and turnover.
Takeaways
- After-call work (ACW) is a critical bottleneck in contact centers, often taking 6.3 minutes for an average 6.5-minute call, leading to high stress and operator turnover.
- The core engineering mission is to use AI to mechanize call summarization and data extraction from audio streams, aiming for a 50% or more reduction in post-processing time.
- The solution is a four-stage low-latency pipeline: Voice Capture, Speech-to-Text (STT) Engine, Generative AI Core, and Customer Data Sync.
- Voice Capture must perform real-time noise filtering, audio normalization, critical channel mapping (separating agent and customer audio), and early-stage Personally Identifiable Information (PII) masking.
- The STT engine requires >90% accuracy, leveraging advanced acoustic modeling, domain-specific dictionaries (e.g., "term life" vs. "term"), inverse text normalization, and auto-punctuation for effective LLM input.
- The Generative AI Core uses an orchestration layer with prompt templates and few-shot learning for structured outputs (e.g., bullet points for inquiry/action), a reasoning layer for intent classification, and a trust layer for token optimization and hallucination checks.
- A Customer Data Sync layer acts as an API gateway to map AI-generated JSON output to CRM fields, incorporating a human verification step before final data commit.
- Implementation resulted in a 50% reduction in ACW (from 6.3 to 3.1 minutes), significantly improved data quality and standardization, and reduced operator cognitive load.
Vocabulary
ACW — After-Call Work; the post-call administrative tasks operators complete, such as note-taking and summarization.
STT — Speech-to-Text; an engine that transcribes spoken audio into written text.
Generative AI — Artificial intelligence models capable of producing new content, such as text summaries or images, from given prompts.
LLM — Large Language Model; a type of generative AI trained on vast amounts of text data to understand and generate human-like language.
PII — Personally Identifiable Information; data that can be used to identify a specific individual, like credit card numbers or passwords.
Hallucinate — In the context of AI, when a model generates information that is plausible-sounding but factually incorrect or fabricated.
Few-shot learning — A machine learning technique where a model can learn a new task from a very small number of examples.
Token optimization — Techniques used to reduce the number of tokens (basic units of text) processed by an LLM, often to manage costs or latency.
Schema mapper — A component that translates data from one structured format (schema) into another, ensuring compatibility between different systems.
Inverse text normalization — The process of converting written numerical or symbolic representations (e.g., "$5,000") back into their standardized numerical format from spoken forms (e.g., "five thousand dollars").
Transcript
Hello everyone, welcome to the AI Engineer 2026 online track. Good morning, good afternoon, good evening, depending upon your time zone. My name is Deepu Singh and I lead the initiatives in the emerging data technologies and AI architecture at Fujitsu North America. And today I will be giving you or maybe provide your deep dive into a very specific but a very highly impactful engineering challenge which we have encountered in our contact center. So with that, let me just quickly share my screen and get this going. All right, so this is the topic of the discussion. WiseOps, a fine, low latency, intelligence extraction from messy audio streams. When we talk about, you know, generative AI, we often focus on clean text inputs but if you go into the real world specifically in the customer service or maybe contact centers specifically, the data does not start as a clean text, right. It starts as messy, quite overlapping, sometimes emotionally charged and it may have multi-channel audio streams also within that, right. And today we will explore the technical architecture which is required to capture that audio, process it with ultra-low latency and use the generative AI to extract structured, actionable business intelligence out of it. So with that, let's get started with this. So here is our roadmap for the next 25 minutes. First, we will stage with what the current challenges are. We have to understand the operational realities and the intense human bottlenecks in the modern contact centers to understand why this engineering matters. Second, we will walk through the solution which is provided. I will break down our high-level architecture into four key components detailing how we move from raw audio to a structured JSON using advanced summarization workflows. Thirdly, we will look at the key outcomes specifically like how these technical implementations they translate into hard ROI and operational impacts. And finally, we will discuss the roadmap ahead and being transparent about the current engineering constraints which we face now and then and where we are taking this technology and the roadmap ahead. So to engineer a great solution, we first need to deeply understand the problem. So let's look at the current state of the contact center operations. So contact centers, they are the front lines of the customer experiences but structurally they are breaking under the pressure. If you look at the industry data, over 50% of the contact centers, they identify hiring, training and productivity as their most critical barriers. Why? Because the job is incredibly difficult. When we analyze the reasons why the operators they leave for the other professions, high stress is the number one factor as part of it. Now operators are expected to handle complex customer emotions, navigate the multiple customer data platforms or the CRM systems involved into it and also document everything perfectly. So if you think about this, this leads to massive retention problem. We are caught in a negative spiral where understaffing leads to higher stress for the remaining operators which leads to height and over and that ultimately implies to more understaffing. So to break this cycle, we just cannot hire more people. We have to fundamentally engineer the stress out of the workflow. And the most clearing inefficiency in this workflow is something called after-call workflow or ACW, which we are going to discuss as part of this slide. According to our baseline studies, the average contact center called it typically last about like 6.5 minutes. However, the average post-processing time where the operator types of the nodes, some rises the call and selects the disposition codes, it takes like almost 6.3 minutes. So this is nearly like one is to one ratio we are talking about. And this also implies operators are spending almost as much time doing administrative data entry as they are actually talking to the customers. Furthermore, because the summarization relies on an individual operator's memory and the writing skills, the data quality is highly inconsistent. Our core engineering mission here was clear like use AI to target the after-call work which is ACW. If we can mechanize the summarization and the data extraction, theoretically we can reduce the post-processing time almost by 50% or even more. And this shifts the enterprise focus from merely just handling the calls to actually analyzing the voice of the customers for the business growth which we are looking out every now and then. So how do we engineer that shift? Let's deep dive into the technical solution and the architecture we built to solve this. So we designed a four stage low latency pipeline to transform the conversational audio into a structured business intelligence with minimal human intervention. And it starts with voice capture which is tapping into the telephony system to extract raw high fidelity audio streams. And that flows into our speech to text engine STT which is responsible for high accuracy transcription. Next is the brain of the system, the generative AI core. This is where we do the heavy lifting of the intent recognition and summarization. And finally, the customer data sync layer which translates those AI insights into API calls to update the customer data or CRM data automatically. Let's look at the engineering under the hood for each of these components in detail. So the first component is the voice capture in AI. The typical rule is garbage in equals garbage out. So if your audio intake is flawed, the LLM will hallucinate later on. We do real time audio intake. So applying noise filters to strip out the back office chatter or anything of those sort which is creating an attenuation is important and normalize the audio level is very, very important for us. Crucially we perform these channel mapping. We absolutely must split the stereo audio to isolate the agent on one channel on the very left and the customer on the other side which is the very right. So if you mix them into a single monotrack, like kind of overlapping with each other, the AI will struggle to figure out who said what and thereby reigning the entire downstream summary. So it's important that we have that channel mapping intact and it separates which who is what and who is saying what right. Finally, we apply a security layer because sometimes the audio streams can contain the credit card numbers or maybe passwords or anything which is personally identifiable information PII. So we utilize buffer management and early stage PII masking technique so that the sensitive data it never hits the LLM memory banks whenever we are moving ahead in the channel. Now next the audio it hits the speech to text engine for generative AI or the LLM to summarize the data effectively or your response effectively. We found that the speech to text the STT accuracy must be above 90%. We utilize advanced acoustic modeling to map the phone maze and filter out any regional dialects. We then apply the language logic utilizing domain specific dictionaries just for example if it's an insurance agent the speech to text engine STT needs to know the difference between a term life and a term right. Both of them are very close to each other but there should be a difference between them. Finally post processing is also vital. We use inverse text normalization and auto-punctuation. For example if a customer says $5,000 the speech to text engine it should must output that into numerical fashion and this numerical formatting it drastically improves the LLM's ability to extract the entities in later point in time. Now we reach the generative AI core we are not just throwing a raw transcript at LLM but we are also asking it to summarize it like we use a highly orchestrated approach in this case. In our orchestration layer we use specific prompt templates our setup showed that if you just ask an LLM to some riser call it output some messy narrative paragraph. So instead we use few short libraries to instruct the LLM to output separate bullet points. One list for customer inquiry and a separate list for operator's action. Then comes the reasoning layer. In the reasoning layer we extract the intent. We provide the LLM with a predefined list of customer called reasons like cancellation or new application or any kind of claim status and instructed to classify the transcript and output the reason why it choose that specific classification. And finally the trust layer where we apply the token optimization to keep the latency low and runs the automated hallucination checks to ensure the generated summary is strict and it is grounded in the transcript. Now the final technical hurdle is getting this beautiful data back into hands of the business. Our API gateway acts as a schema mapper. It takes the JSON output from the LLM and maps the field like customer intent or resolution status and directly to some corresponding fields based on the company's CRM system or any customer data which we have laid out via some REST APIs. So we don't remove the human entirely in this. We use a verification step in between the operator. It sees the AI generated summary auto-populated on the skin. They do a quick visual field validation. Make some minor edits if necessary and then just click the confirm. Samultaneously this structured data it flows into our business intelligence models aggregating the voice of the customer data for management dashboards and automatically you know flagging the candidates for new FAQ data entries. Now to tie the architecture together this is the linear workflow logic of the data pipeline. We take the raw transcript complete with the time indexing confidence scoring and denoising. We pass it through the speaker separation because we split the stereo channels in the step one. We can easily stitch the dialog together logically like customer said x agent said y. Then we move to the context deduction where the lm spots the entities like account numbers or product name or customer name run the sentiment analysis and recognizes the intent. The final state is the structured output so instead of a wall of text the system it outputs like clear and a clean JSON schema matching some predefined customer data or maybe CRM templates which we have in the enterprise and then categorize neatly into bullet points and this strict formatting is what turns an unstructured conversation into a database ready asset. So what happens when we deploy this architecture in a real contact center environment. Let's look at the outcomes. The operational impacts out of the implementation were quite immediate and highly measurable. Look at the ECW time. Under the manual operation after average after call work was 6.3 minutes powered by our AI workflow that dropped to like 3.1 minutes which is like almost 50% reduction in the processing time. If you calculate that across 500 seats handling thousands of calls a day you are looking at a massive operational saving the equivalent of almost reclaiming dozens of full time headcounts purely from efficiency standpoint. The next aspect is of data entry quality it moved from a highly subjective and a variable to highly standardized and uniform output. The inquiry categorization or the call reason tagging it moved from being dependent on an operator mode or memory to being strictly logic based resulting in a highly consistent voice of the customer data set for the management. And ultimately by removing these repetitive administrative burden of typing out notes we reduce the cognitive load on the operators thereby stabilizing the operations and directly combating stress which is linked with the staff and this ultimately reduce the turnover we identified at the very beginning of our discussion. While the results were fantastic the engineering work is never done and let's talk about the constraints we face in our roadmap for the future. We are currently navigating three main constraints first is the source to text accuracy, STT accuracy. The entire generative AI summary it relies on the transcript so if the STT engine fails to pick up heavy accents or poor audio quality the LLM has nothing to work with so the STT optimization is a continuous battle for us. Second is the initial setup cost while the long term ROI is massive. The initial consumption of the API tokens especially running complex LLM reasonings on long 20 minute transcripts can be costly and especially during the initial scaling those are tough ones and we are constantly working on token optimization techniques to bring this number down. The third is the security and compliance handling PII any kind of sensitive information in audio streams is very complex and ensuring robust masking before the data hits Claude endpoint is a strict requirement and because of this we add some layers and ultimately reduces the latency and also adds some overhead from architectural standpoint so we are still figuring it out how we can reduce those extra layers and make it much more robust component-wise. So to address these constraints and push the boundaries of what's possible our roadmap is currently broken down into three phases. Phase 1 it focuses on explainable AI. We want to move beyond just summarizing calls to actually you know coaching the operators. We are engineering the systems to analyze the audio post call and provide operators with instant private feedback on their either soft skills or empathy level or any kind of accuracy from a information standpoint. The phase 2 it targets the predictive staffing by taking the massive amount of categorized intent data we are now capturing. We can feed the same exact data intent data into a time series analytics and this will allow the workforce management to accurately forecast the call volumes spikes based on those specific topics and thereby optimizing the shift scheduling. The phase 3 is perhaps the most important for human well-being combating the customer harassment. Contact center agents they mostly face an increasingly amount of verbal abuse and we are developing a low latency sentiment and acoustic analysis that can detect when a customer becomes abusive. Ultimately the system can you know trigger some triggers I mean analysts or something which is important from notification standpoint to supervisor or anyone who is there from the management upper management standpoint or maybe seamlessly transfer the call to an AI voice agent just to protect the human operator mental health in case of these tough conversations. All right so by applying these rigorous engineering techniques to the messy audio data we can definitely transform the contact centers from cause centers of like high stress into a highly efficient and intelligence gathering engines that protect their work forces that's the whole idea of having this right. Thank you so much for your time and listening to me. I have included my QR code which I will just flash in here that you can grab me over LinkedIn and please feel free to connect if you would like to discuss the architecture or any prompt engineering strategies related to the discussion which we had in here and happy to connect and make things work things out between us and we can have more conversations. So thank you so much for listening to me and have a good one. Bye.