How AI Doubled the Engagement Rates of a Virtual Event Platform
Service:
Custom AI Development, Machine Learning Engineering
Client:
Social27
Industry:
Virtual Events & SaaS
Location:
USA
Client Website:
Social27 is a Fortune 500-trusted virtual event platform used by Microsoft, the United Nations, Capgemini, and United Airlines. It runs large-scale virtual, hybrid, and in-person events globally. PrimeSens designed and built a proprietary AI-powered Recommendation Engine that personalised session discovery, peer matching, and lead targeting for millions of attendees across the platform.
Challenges
Why do virtual event platforms struggle with attendee engagement at scale?
Virtual events have a problem nobody likes to say out loud. Attendees log in, scroll through a wall of sessions, feel immediately overwhelmed, and disengage within minutes. There is no helpful conference volunteer to point them toward the right room. There is no accidental hallway conversation that sparks a connection. The serendipity of physical events simply does not transfer to a digital screen by default.
For a platform the size of Social27, serving Fortune 500 clients who host thousands of attendees per event, this was not a minor UX irritation. It was a core business threat. Sponsors were paying for visibility. Event owners were promising ROI. Attendees were leaving sessions early, skipping networking entirely, and returning engagement metrics that told a story nobody wanted to read.
How does poor session discoverability hurt event ROI for sponsors and organisers?
Sponsors invest significant budget into virtual events expecting a return. That return depends on one thing: the right people seeing their content, engaging with their brand, and being identified as genuine leads. When attendee navigation is random and unguided, sponsor sessions get low attendance not because the content is poor but because the wrong people are finding it, or nobody is finding it at all.
Social27's event owners needed to demonstrate concrete ROI to their Fortune 500 clients. That meant session watch time, booth interactions, and qualified lead data. The platform had the audience. It had the content. What it lacked was an intelligent layer that could connect the two with precision.
What makes building a personalisation engine for live events technically difficult?
Most recommendation systems operate on historical data. Netflix knows what you watched last week. Spotify knows your listening history. But live events are different. An attendee's behaviour at 9am on Day 1 needs to influence what they see by 10am. Data is arriving in real time. Preferences are shifting. And the system needs to process millions of records across thousands of concurrent users without lag, without failure, and without serving the same irrelevant content twice.
This is not a problem you solve with a standard off-the-shelf recommendation library. The ETL (Extract, Transform, Load) pipelines need to handle volume. The ML models need to be fast enough to operate meaningfully within a live event window. And the entire infrastructure needs to be distributed and fault-tolerant. Social27 needed an engine built for the specific, high-stakes, time-sensitive nature of live virtual events. That kind of system does not exist out of the box.
Why is lead quality a bigger challenge than lead volume in virtual events?
Any event platform can hand a sponsor a list of badge scans. That is not valuable anymore. What sponsors, particularly enterprise-level ones, want to know is: who among these thousands of attendees is actually worth a follow-up call? Who has a high spending propensity? Who is actively in a buying cycle versus just attending because their company paid for it?
Without intelligent lead matching, event owners were delivering raw data dumps and calling them leads. Sponsors were frustrated. Renewal rates suffered. The platform needed to shift from delivering volume to delivering quality, and that meant building a lead scoring and matching engine that could evaluate behavioural signals in real time and surface the highest-value prospects with confidence.
Solution
Solution
Building a Real-Time AI Recommendation Engine for Live Virtual Events
PrimeSens was engaged to design, build, and maintain Social27's proprietary Recommendation Engine from the ground up. This was not a configuration project. It was full custom AI development, engineered specifically for the demands of large-scale live events.
The engine operates across four core recommendation surfaces: personalised sessions, peer matches, round table suggestions, and relevant advertising placements. Each of these surfaces is driven by a shared understanding of the attendee built from behavioural signals captured and processed throughout the event in real time.
When an attendee joins a session, that signal feeds the model. When they visit a booth, that registers. When they connect with a peer, that updates their interest graph. The system is always learning, always adjusting, and always getting more accurate as the event progresses.
Proprietary Behavioural Analysis and Interest Profiling
The foundation of the engine is a behavioural analysis layer that tracks and interprets attendee actions throughout the event lifecycle. This layer does not rely on self-reported preferences or pre-event survey data, though that data is incorporated where available.
Instead, it builds a dynamic interest profile for each attendee based on what they actually do. Sessions watched and for how long. Topics engaged with. Connections made. Booths visited. Questions asked in live Q&A. Each action is weighted and fed into a continuously updating ML model that shapes every subsequent recommendation the attendee receives.
This matters because stated preferences and actual behaviour diverge constantly. Someone may tick "marketing" as their primary interest in a pre-event form but spend most of their time in product and engineering sessions. The engine picks that up immediately and adjusts. It follows behaviour, not declarations.
Distributed Spark Cluster Architecture for Million-Record ETL Processing
The engineering challenge at the core of this project was scale. Social27 events can host thousands of simultaneous attendees generating continuous interaction data. That data needs to move through ETL pipelines, be processed by ML models, and return personalised recommendations fast enough to feel instant.
PrimeSens architected the data infrastructure around a distributed Apache Spark cluster. This allowed the ETL processes handling millions of records per event to run in parallel, with the compute load distributed across the cluster rather than bottlenecked at a single point. ML model inference runs within this distributed environment, meaning the system scales horizontally as event size grows.
The result is an engine that does not slow down when events get bigger. It handles the load, processes the data, and returns recommendations in real time regardless of the number of concurrent users on the platform.
Lead Matching Engine for Sponsors, Businesses, and Event Owners
Separate from but integrated with the attendee recommendation layer, PrimeSens built a Lead Matching Engine designed specifically for sponsors, exhibitors, and event owners attending the event.
This engine evaluates behavioural signals, session attendance patterns, booth engagement depth, and inferred interests to score attendees on their spending propensity and match them to relevant sponsor profiles. A cybersecurity sponsor, for example, does not need a list of every person who visited their booth. They need the 47 attendees who spent more than eight minutes at the booth, watched the keynote on enterprise security, and hold a VP or C-level title.
The Lead Matching Engine surfaces that list automatically. Sponsors receive qualified leads, not raw lists. Event owners can demonstrate measurable ROI. And the platform's value proposition to Fortune 500 clients becomes defensible and repeatable.
Continuous Model Maintenance and Performance Optimisation
Building the engine was one part of PrimeSens's engagement. Maintaining it was another. ML models degrade over time as attendee behaviour patterns shift and as the platform grows into new use cases and event formats. PrimeSens remained responsible for ongoing model monitoring, retraining, and performance tuning post-launch.
This included monitoring recommendation click-through rates and session watch completion as proxy metrics for model accuracy, running A/B tests on recommendation logic, and iterating on the feature engineering that feeds the models. The engagement was not a handoff. It was a sustained technical partnership built to keep the engine performing at its peak as Social27's client base and event complexity continued to grow.
Results
Result
63% longer session watch times across events using the Recommendation Engine
2x increase in platform-wide attendee engagement rates
41% higher lead conversion for sponsors using the Lead Matching Engine
Event owners reported that sponsors renewed contracts at a significantly higher rate after AI-powered lead data replaced raw attendee lists
Fortune 500 clients including Microsoft and United Airlines saw measurably higher attendee satisfaction scores attributed to personalised session discovery
Average attendee interaction depth, measured by unique sessions attended, booth visits, and peer connections made, increased by over 55% compared to pre-engine baseline data
The distributed Spark architecture handled peak loads of over 2 million records per event without performance degradation
Round table and peer match recommendations drove a 38% increase in attendee-initiated networking connections per event
Social27's Net Promoter Score from event owners improved by 29 points within two event cycles following full deployment of the engine
But the bigger win was strategic. Clandestine's fraud team shifted from firefighting to pattern recognition — identifying emerging threats before they scaled, tuning the system proactively, and finally operating ahead of the curve. Fraud prevention transformed from a cost center into a competitive advantage.
The engagement metrics Social27 achieved are not a coincidence. They are the direct output of an intelligent system built to replace guesswork with precision, and passive browsing with genuinely personalised discovery. If your platform, product, or business is sitting on behavioural data that is not yet working for you, that is the exact problem PrimeSens was built to solve. The gap between where you are and what your data could be doing for you is shorter than you think.
Frequently Asked Questions
FAQ
What is an AI recommendation engine for virtual events?
An AI recommendation engine for virtual events is a machine learning system that analyses attendee behaviour, preferences, and real-time interactions to surface personalised session suggestions, networking matches, and relevant content. It works by processing signals like session attendance, booth visits, and click behaviour to build a dynamic interest profile for each attendee, then matches that profile against available content and peers. The goal is to replace random browsing with guided, personalised discovery that keeps attendees engaged longer.
How does AI personalisation increase event engagement rates?
AI personalisation increases event engagement by reducing friction in content discovery. When attendees are shown sessions and connections that are directly relevant to their interests and behaviour, they stay on the platform longer, attend more sessions, and interact more deeply with sponsors and peers. Generic event navigation produces passive attendees. Personalised navigation produces active participants. The difference in engagement metrics between the two is consistently significant across platforms that have implemented recommendation systems.
What is a lead matching engine and how does it work in virtual events?
A lead matching engine is an AI system that evaluates attendee behaviour during an event and scores individuals on their likelihood of being a high-value prospect for a specific sponsor or business. It works by analysing signals like session attendance, booth engagement duration, interaction patterns, and inferred professional interests, then matching those signals against sponsor profiles and product relevance. The output is a ranked list of high-propensity leads, not a raw list of badge scans.
What technology stack is used to build a real-time event recommendation system?
A production-grade real-time recommendation system for virtual events typically includes:
Data pipeline layer: Apache Spark or Apache Flink for distributed ETL processing at scale
ML model layer: Collaborative filtering, content-based filtering, or hybrid models built with frameworks like TensorFlow, PyTorch, or Scikit-learn
Feature store: A real-time feature store such as Feast or Tecton to serve pre-computed features at low latency
Serving infrastructure: REST APIs or gRPC endpoints for low-latency model inference
Orchestration: Apache Airflow or Prefect for pipeline scheduling and monitoring
Cloud infrastructure: AWS, GCP, or Azure for scalable compute and storage
The exact stack depends on event scale, latency requirements, and existing platform architecture.
How long does it take to build a custom AI recommendation engine?
A custom AI recommendation engine for a production SaaS platform typically takes between 3 and 6 months to build from requirements to initial deployment, depending on data availability, infrastructure complexity, and the number of recommendation surfaces required. Simpler single-surface engines can be scoped tighter. Engines that handle multiple surfaces (sessions, ads, peer matching, and lead scoring simultaneously) and need to operate in real time during live events require more architecture investment and testing cycles before they are production-ready.
What is the difference between a custom AI engine and an off-the-shelf recommendation tool?
An off-the-shelf recommendation tool is a pre-built system configured to your data. A custom AI engine is designed, trained, and optimised specifically for your platform's use case, data structure, and performance requirements. Off-the-shelf tools work for general cases. They struggle with highly specific contexts, like the real-time, time-bound, multi-surface nature of live virtual events, where generic models underperform because they were never trained to understand the unique behaviour patterns of event attendees. Custom engines cost more upfront but consistently outperform generic tools in accuracy, latency, and business impact.
Can AI improve virtual event ROI for sponsors?
Yes. AI improves virtual event ROI for sponsors by replacing volume-based lead delivery with quality-based lead delivery. Rather than receiving a list of everyone who visited their booth, sponsors receive a shortlist of attendees who demonstrated genuine buying signals throughout the event. This leads to higher sales conversion rates from event-sourced leads, stronger sponsor satisfaction scores, and better contract renewal rates for the event platform. The shift from raw data to intelligent lead scoring is one of the clearest ROI levers available in the virtual events space.
How does a distributed Spark architecture support large-scale event data processing?
Apache Spark processes large datasets by splitting jobs across a cluster of machines and running them in parallel. For virtual events handling millions of behavioural records from thousands of simultaneous attendees, this distributed approach prevents the bottlenecks that single-node processing systems hit under peak load. Spark's in-memory processing also makes it significantly faster than traditional disk-based MapReduce approaches, which matters when recommendation freshness during a live event is measured in seconds, not minutes.
What kinds of businesses benefit most from custom AI recommendation systems?
Businesses that benefit most from custom AI recommendation systems share a few common traits: they have large user bases interacting with large content or product catalogues, they have meaningful behavioural data they are not yet using, and they have a clear business metric, such as engagement, conversion, or retention, that personalisation can directly move. This includes SaaS platforms, e-commerce businesses, media and content companies, online marketplaces, and enterprise event technology platforms. If you have the data and a meaningful metric to improve, a custom recommendation system is almost always worth evaluating.
How do I know if my platform needs AI personalisation or just better UX?
Better UX solves navigation problems. AI personalisation solves relevance problems. If your users can find content easily but still disengage, the problem is relevance, not navigation. If your users struggle to find content because the catalogue is too large for manual browsing to work, the problem is both. A useful diagnostic: look at the breadth of content your users engage with versus the breadth available. If users are only engaging with a narrow slice of what is on offer, personalisation is likely to unlock significant engagement gains. If they are engaging broadly but drop off quickly, the issue may be content quality or UX friction instead.
