This interview is with ANSHUMAN GUHA, Staff Engineer Data Scientist at Freshworks.
ANSHUMAN GUHA, Staff Engineer Data Scientist, Freshworks
Can you introduce yourself and highlight your key achievements in the field of machine learning and data science?
Hello, I’m Anshuman Guha, a dedicated Machine Learning Engineer with over seven years of experience in deploying advanced AI solutions, enhancing NLP systems, and leading high-impact projects. My journey has been centered around translating research into scalable engineering products while optimizing machine learning models.
At Freshworks, where I currently serve as a Staff Engineer – Data Scientist, I’ve worked on projects that significantly enhanced our FreshChat and Freshdesk CRM products. One of my notable achievements was leading the deployment of multilingual and multimodal features, which resulted in the release of Freddy AI Copilot. This initiative improved agent interaction quality across languages and scaled to 500k monthly interactions.
In my previous role at Capital One, I spearheaded credit card fraud model deployments with a notable net positive value of $500 million, drastically reducing deployment turnaround times. Additionally, I played a crucial role in creating an innovative ML model scoring deployment pipeline that significantly minimized production errors.
I hold a Master’s degree in Computer Science from Johns Hopkins University, which underpins my technical expertise. Throughout my career, I’ve been privileged to present and speak at prominent conferences such as DSS-SF and AIAI-Austin, sharing insights on leveraging AI for real-time quality control and customer satisfaction enhancement.
For me, pushing the boundaries of machine learning and ensuring the technology aligns with ethical standards like fairness and correctness has always been a priority. I’m passionate about continuing to contribute to this dynamic field by driving innovation and fostering collaborative teamwork.
How did your journey in machine learning and data science begin, and what pivotal moments led you to your current role at Freshworks?
My journey in machine learning and data science began with my master’s in Computer Science from Johns Hopkins University, where I developed a solid foundation in the principles underlying these fields.
A pivotal moment came during my tenure at SparkCognition, where I led the development of neural network models for failure detection in industrial IoT, culminating in a presentation at the OSDC West Conference in 2017.
My experience grew significantly at Capital One, where I spearheaded credit-card fraud models and reduced ML errors in production. The patent I filed for an ML model scoring deployment pipeline was a notable highlight, further solidifying my expertise in MLOps.
Transitioning to Freshworks as a Staff Engineer in AI Labs, I applied my skills to develop large language models and launched impactful features like the multilingual capabilities in FreshChat and the Freddy AI Copilot.
Leading projects like the Proactive Quality Coach and developing advanced RAG bots were crucial in shaping my current role. Presenting research advancements at conferences like DSS-SF and AIAI-Austin and receiving press coverage were significant milestones in my professional journey.
These experiences, coupled with speaking engagements and published articles, were instrumental in propelling me to my current position at Freshworks, where I continue to innovate and lead projects that transform AI research into practical applications.
You’ve developed three patented or patent-pending product features. Can you share the story behind one of these innovations and how it addresses a real-world problem?
One of the innovations I’m particularly proud of is the “Proactive Quality Coach” (PQC) feature I spearheaded at Freshworks. This tool addresses the real-world problem of improving agent communication in customer service by providing real-time write-assist capabilities using Generative AI.
By enhancing grammar, relevance, and tone across multiple languages, PQC supports over 20,000 agents and handles around 500,000 interactions monthly. This not only elevates customer satisfaction but also streamlines agent efficiency, making it an essential tool for global businesses to maintain high-quality interactions.
In your experience leading the development of multilingual features for AI-powered chatbots at Freshworks, what were the biggest challenges you faced, and how did you overcome them?
In my experience leading the development of multilingual features for AI-powered chatbots at Freshworks, one of the biggest challenges was ensuring the accuracy and relevance of language translations across multiple languages. Handling nuances and cultural differences while maintaining meaningful interactions was complex.
To overcome this, I focused on instruction tuning of LLMs with techniques like reinforcement learning from human feedback (RLHF). By extending the CoEDIT research to 15 languages using LoRA, I improved various language features, ensuring minimal edits and preserving native accents. I also implemented custom Hugging Face logit processors to prevent nuisance edits and reduce biases.
Another challenge was scaling the system to handle millions of interactions monthly. I addressed this by architecting a Fast API-based pipeline to reduce latency significantly. This involved using task-specific models, quantization, and efficient payload management, ultimately achieving an 80% reduction in latency for LoRA fine-tuned models.
Through these strategies, I successfully deployed impactful multilingual features that enhanced our chatbot’s global usability and effectiveness.
Can you share a specific instance where your expertise in MLOps significantly improved a project’s outcome, and what lessons can other data scientists learn from this experience?
One specific instance where my expertise in MLOps significantly improved a project’s outcome was during my time at Freshworks. I led the development of the Proactive Quality Coach, which scaled to 500,000 interactions monthly, enhancing multilingual support and improving agent performance through a Gen-AI-based real-time write-assist product.
I architected a Fast API-based pipeline that reduced latency by 80% for our fine-tuned 8B LLMs. This involved creating task-specific precursor models using small LMs, quantization, and payload management.
The lesson for other data scientists is the crucial role of efficient MLOps practices, such as automating pipelines and optimizing model deployment processes, which can drive significant improvements in performance and scalability. Implementing robust testing and reducing latency are key factors in achieving real-time application success.
As someone who has presented at multiple conferences, what advice would you give to data scientists looking to effectively communicate complex ML concepts to diverse audiences?
As someone who has presented at multiple conferences, I would recommend a few strategies for data scientists aiming to effectively communicate complex ML concepts to diverse audiences:
1. **Know Your Audience:** Tailor your presentation to the knowledge level and interests of your audience. Use simpler language if they are not experts in the field, and incorporate technical depth if they are peers.
2. **Use Relatable Analogies:** Translate complex ideas into relatable concepts using analogies or stories that the audience can easily understand.
3. **Visual Aids:** Leverage visuals like charts, diagrams, and animations to illustrate concepts. This can make abstract ideas more tangible and easier to grasp.
4. **Highlight Real-world Applications:** Focus on how the technology impacts real-world scenarios. Demonstrating practical applications can make the data science concepts more relevant and engaging.
5. **Engage with Questions:** Encourage questions throughout the presentation to clarify doubts and engage the audience. This interaction keeps the session dynamic and inclusive.
6. **Simplify Complex Concepts:** Break down complex algorithms or models into simpler components and explain each part step by step.
7. **Storytelling:** Craft your presentation as a narrative. A compelling story can keep your audience engaged and make your points more memorable.
8. **Practice and Feedback:** Rehearse your presentation multiple times and seek feedback from colleagues or friends to refine your delivery and content.
These strategies have helped me convey intricate topics effectively at conferences like DSS-SF and AIAI-Austin, ensuring my message resonates with diverse audiences.
In your work with IoT devices at SparkCognition, how did you approach the challenge of modeling ‘normal’ behavior for such diverse systems, and what insights can be applied to other domains?
In my work with IoT devices at SparkCognition, I approached the challenge of modeling ‘normal’ behavior by developing neural network models like Multi-Channel CNNs and sliding-window autoencoders. These models were crucial in understanding ambient conditions, addressing equipment aging, and dealing with the scarcity of failure data. By conceptualizing and building these models, I was able to detect failures in the industrial IoT domain effectively.
The insights from this experience can be applied to other domains by emphasizing the importance of understanding the baseline or ‘normal’ conditions for any system. Addressing data scarcity and continuously adapting models to account for changes in environments are crucial strategies that can be universally applied to enhance predictive modeling in diverse domains.
Looking at the rapidly evolving field of AI and ML, what area do you believe holds the most promise for impactful innovations in the next five years, and why?
In the rapidly evolving field of AI and ML, I believe that advancements in Large Language Models (LLMs), specifically in the realm of natural language processing and real-time quality control, hold the most promise for impactful innovations in the next five years.
My experience with deploying generative AI solutions and enhancing NLP systems has shown me the tremendous potential these technologies have in transforming customer service and operational efficiency.
Multimodal features and multilingual capabilities are increasingly important as global communication expands. The development of advanced retrieval-augmented generation (RAG) systems to improve search precision and relevance could significantly enhance user interactions across industries.
Moreover, focusing on bias prevention, safety, fairness, and correctness of LLMs at scale is vital. As I’ve worked on improving these aspects, I see great promise in instruction tuning and the use of reinforcement learning from human feedback (RLHF) to make AI models more aligned with human values.
This focus will not only improve customer satisfaction and operational effectiveness but also help in managing diverse data interactions, leading to more inclusive and efficient AI systems.