As a genuinely AI-native platform, artificial intelligence is the foundational infrastructure at Theta Lake, not an add-on feature. Our first corporate hire was a Chief Data Scientist, our compliance classifiers leveraged AI from inception, and our architecture is underpinned by proprietary patents dating back to 2018.
While AI has long been at our core for improving communication compliance, machine learning (ML) is the functional engine that makes that possible. And as AI expands to govern a brand-new landscape of AI-driven tools and workplace behaviors, it relies on increasingly sophisticated ML models to understand the nuance of human interaction.
Recognized as a Visionary in the Gartner® Magic Quadrant™, and backed by independently verified ISO 42001 and CSA Star for AI Level 2 certifications, we pride ourselves on building a compliance platform that illustrates the difference between foundational AI and superficial marketing claims. Below, Rohit Jain, Distinguished Engineer at Theta Lake, explains exactly how our machine learning architecture has evolved to meet these challenges.
Foundation: Designing for a Multimodal World
From day one, Theta Lake was built for a multimodal world. We recognized that legacy lexicon-based systems designed for email were insufficient for the nuance and context of modern communication over channels like video and chat. Lexicons cannot generalize beyond syntax into semantics; they suffer from incredibly high false positive rates because they lack context awareness. We knew we needed to solve the twin challenges of semantic generalization and context awareness—a necessity that has only grown more urgent as evolving communication modalities rapidly reshape language itself.
Adding Robustness: Beyond Keywords
Our first major hurdle was moving beyond rigid keyword detection. Because our customers were accustomed to traditional lexicons, we wanted to utilize their existing work and immediately improve its accuracy without adding to their operational burden. We engineered proprietary IP (U.S. Patent No. 12,045,561) to handle lookalikes and soundalikes, ensuring our system remains robust against misspellings, OCR errors, and semantic variations. This advanced our generalization capabilities: we don’t just match words; we understand intent.
As we developed more text-matching use cases—such as disclaimer detection and transcript matching—we developed IP for fuzzy matching across larger sections of text, alongside corresponding error models for OCR and transcription (U.S. Patent No. 12,265,563), extending our pattern-matching footprint even further.
Improving Data Inputs
The development of our pattern-matching system was directly tied to the error patterns we observed in transcription and OCR (optical character recognition) inputs. Realizing that our models required the highest-quality data inputs, we rigorously tested multiple third-party OCR and transcription engines under various audio/video conditions, speakers, and accents. This extensive research resulted in our bespoke, proprietary transcription test suite, which we run periodically to evaluate, benchmark, and select the highest-performing vendors.
The Power of Ensembles
While some of our classifiers still utilize simpler lexicons, keywords and enhanced pattern-matching techniques, others require cutting-edge machine learning algorithms.
We began with word embeddings and developed a method to selectively combine them to detect specific sets of concepts and phrases associated with a given classifier (U.S. Patent No. 12,265,563). While successful, our initial attempts to fully automate this specific improvement process fell short of expectations—so we shifted focus to emerging sentence embeddings. After testing multiple variations across our classifiers, we selected the embeddings that demonstrated the optimal balance of sensitivity and specificity on our datasets. As we matured these embedding workflows, we identified and addressed their unique failure modes, while adding capabilities to cleanly ingest and leverage non-text elements like emojis within our training data and use them effectively in our classifiers.
We used these embeddings as features, feeding them into ensembles of traditional machine learning methods ranging from simple Naive Bayes to neural networks, tree-based methods, boosting, GBMs, and KNeighbors. This ensembling process is entirely data- and metrics-driven. Initially managed manually, this workflow became semi-automated and is now almost completely automated.
As we built more complex classifiers, we recognized that the embedding-based approach wasn’t always the absolute best solution. We found that discriminative and generative language models, fine-tuned to our intelligently selected and accurately labeled data, could outperform embedding-based ensembles in specific cases, so we introduced them selectively where the accuracy gains justified the substantial performance overhead.
Today, every classifier we build is a custom-tuned ensemble, evolving from simple approaches to complex algorithmic mixes. This iterative, data-driven methodology allows us to constantly refine performance, introducing new techniques and algorithms precisely where they matter most.
Similarly, for visual data, we started with state-of-the-art object detection for specific categories. Over time, as these categories expanded, we integrated image classification methods for select use cases. Just as with our text workflows, our choice of techniques and visual algorithm selection remains entirely data-driven. An interesting use of these capabilities is highlighted in our methods for detecting applications shared on screen (U.S. Patent No. 12,464,032 B2).
In Part 2 of this series, we move from foundational model building to scaling—exploring how we monitor these living systems in production, expand globally, and use AI to solve compliance’s toughest “needle in a haystack” data problems.
