Artificial Intelligence (AI)
Discuss current events in AI and technological innovations with Intel® employees
525 Discussions

Revolutionizing Email Generation: SiteMana's Tailored LLM Built on Intel® Technology

0 1 3,168

The vast majority of visitors a website gets are anonymous - a fact that marketers and online vendors know only too well. These visitors come and go without giving their email address or any other information about who they are. For businesses, this traffic represents huge untapped potential, tantalisingly out of reach.

Intel® Liftoff member SiteMana is revolutionizing the industry by providing businesses with a secure and legal method to engage with anonymous traffic, all while respecting individual privacy and without uncovering personal identities. Through advanced machine learning, SiteMana identifies visitors with high buyer intent and retargets them with personalized messaging. In SiteMana's own words, they enable their clients to “master the conversion process,” transforming anonymous visitors into lifetime customers.


SiteMana Joins the Intel® Liftoff Hackathon


SiteMana’s vision is to optimize abandoned cart recovery and generate special deals tailored to user browsing interest. To achieve this, they sought to develop a finely tuned natural language processing model, aiming to create a solution that could complement or even surpass the capabilities of OpenAI’s GPT-4 API, which they had previously relied upon. Despite the recognized strengths of the GPT-4 API, SiteMana experienced challenges related to speed and throttling limitations. After joining the Intel® Liftoff Hackathon, they embarked on a journey to overcome these obstacles.

"Thanks to the remarkable capabilities of Intel's GPU, Dolly 2.0, and OpenLlama provided during the hackathon, we, at SiteMana, constructed an LLM model akin to cutting-edge chatbots. This model underwent meticulous fine-tuning to craft individualized emails, and the deployment and testing phases proceeded with unexpected ease. With the immense potential of Intel's GPU driving us forward, SiteMana harnessed Dolly 2.0 and OpenLlama during the hackathon to tailor an LLM model to our specific needs. This model, dedicated to composing personalized emails, was seamlessly integrated and tested, surpassing our initial expectations," as articulated by Peter Ma, the founder of SiteMana.


The Workload


image (7).pngInitially, SiteMana explored various models, including loading up Dolly-v2-7b, and benchmarking with an openLlama. During this process, the Intel® Data Center GPU Max 1100 played a pivotal role. With its 48 GB of HBM2e memory and 1228.8 GB/s graphics memory bandwidth, this GPU provided the necessary performance to enable seamless processing.

The collaboration of SiteMana's vision with Intel's robust hardware led to the development of an API that facilitated the real-time generation of personalized content through natural language processing algorithms.

SiteMana extended their hardware optimization beyond GPUs by utilizing the Intel Developer Cloud for Large Language Model (LLM) inference on both XPUs and CPUs. This integration allowed them to deploy the 4th Generation Intel® Xeon Scalable Processors in conjunction with the Intel® Data Center GPU Max 1100, enhancing the efficiency of their algorithms. For the specific task of predicting a user's purchasing intent, they employed Intel oneAPI. This well-rounded approach to hardware utilization demonstrates their focus on leveraging the right tools to meet their technical objectives.




The Results


By leveraging Intel's advanced hardware and software, SiteMana successfully deployed a robust language model, ManaLLM, specifically tailored for email generation. This innovation led to enhanced email retargeting capabilities, allowing for improved service delivery. For SiteMana, this development represents not only a technical advancement but also a promising avenue for future revenue growth.


image (8).png


These results showcase the efficiency and performance gains achievable through the targeted use of hardware. SiteMana’s Peter Ma sees even more potential, expecting these numbers to improve further with the incorporation of multiple GPUs and the utilization of batch processing techniques.

Their next priority will be fine-tuning ManaLLM to outperform GPT-4, their current benchmark. This ambitious goal will involve incorporating customer-specific data and optimizing training algorithms. By doing so, ManaLLM will provide an enhanced service for SiteMana's customers, side-stepping the throttling issues they faced previously with GPT-4. This represents a major step forward in SiteMana's mission to set a new standard in AI-powered email retargeting.


Intel® Liftoff: The Tech Startup Accelerator Bringing Cutting-Edge AI Solutions to Urgent Business Challenges


SiteMana’s results are a testament to the power of Intel’s accelerated computing solutions to help LLM startups like SiteMana bring their most ambitious visions to life. And we’re looking for more innovators, visionaries, movers and shakers to join Intel® Liftoff for Startups. Apply to the tech startup accelerator program to realize the full potential of your technology.

About the Author
I'm a proud team member of the Intel® Liftoff for Startups, an innovative, free virtual program dedicated to accelerating the growth of early-stage AI startups.
1 Comment

Thanks for the coverage!