As financial institutions drown in communication data, the real battle in RegTech is no longer building AI models, but keeping them accurate, scalable and reliable over time.
In the second instalment of its machine learning series, compliance platform Theta Lake has set out how continuous monitoring, disciplined data curation and pragmatic use of Large Language Models (LLMs) turn vast data streams into actionable compliance intelligence.
Theta Lake recently discussed the evolution of machine learning and scaling machine learning for compliance.
Central to the firm’s approach is a rejection of hyper-customised models built for each customer site, a strategy it argues leads to unmaintainable codebases that drift out of date and are eventually abandoned.
Instead, Theta Lake standardises its core models across all customers, allowing only risk thresholds and minor operational parameters to be fine-tuned. Models are updated dynamically based on customer feedback on false positives and negatives, alongside internal tracking of hit rates and model drift, with all performance metrics monitored through central dashboards.
The stakes are high because non-compliant or unprofessional behaviour typically accounts for less than 0.1% of corporate communication flows, creating a needle-in-a-haystack problem that makes classifiers extraordinarily hard to train.
Any system designed to catch all bad behaviour will inevitably flag some benign activity, so Theta Lake has spent years engineering a proprietary matrix of metrics to minimise false positives without sacrificing recall.
Data curation is another pillar. The company’s patented Smart Labeling technology (U.S. Patent No. 12,664,235) intelligently selects training data from vast datasets, runs automated label checks and surfaces potentially mislabelled examples for human verification. This allows classifiers to actively sample under-represented data, cutting both dataset size and compute time.
Scaling globally brought fresh challenges, particularly accurate language identification across short text snippets and code-switching between languages mid-conversation.
The firm built custom in-house language detection test suites and found it also needed to quantify transcription reliability and detect silence in noisy audio to prevent misclassifications. Fine-tuned smaller language models now generalise its core English classifiers across European and CJK languages.
On the generative front, Theta Lake has been an early access beta partner with vendors including Anthropic, deploying its first LLM-powered feature, chat summarisation, three years ago. Yet its testing shows custom-tuned smaller models often outperform LLMs for specific classification tasks, being more cost-effective and easier to debug. Vision Language Models have nonetheless been integrated into its image pipelines, competing with traditional object detection approaches.
The firm’s Alert Confirmation Analysis layer evaluates entire records rather than isolated sentences, delivering risk scores and rationales that let compliance teams automatically prioritise review queues, ensuring the highest-risk communications are audited first.
Read Theta Lake’s full post here.
Copyright © 2026 RegTech Analyst
Copyright © 2018 RegTech Analyst
