📢 Introducing Janus - our new family of customizable AI models for US financial disclosures. This comes with our new bespoke ML architecture, which combines LLMs with cutting-edge few-shot classification techniques.
❓ Why?
Disclosure identification and verification are some of the most important things a compliance officer looks at. Disclosures come in all forms, like:
1. Mandatory disclosures like "XYZ is a technology company and not a bank", the banking partner providing FDIC insurance, etc.
2. Product-specific ones like substantiation for APYs, terms and conditions for offers and cashback, etc.
3. Domain- and company-specific disclosures like SEC/FINRA disclosures, performance returns, etc.
Disclosures themselves do not necessarily protect against deceptive claims, but inaccurate or missing disclosures have historically resulted in enforcement actions (cue Chime and SoFi).
While building Sei AI (backed by YC & PayPal) AI-powered regulatory compliance platform, we quickly realized the need for a model that can accurately extract, validate, and provide suggestions to add the correct disclosures.
🔨 How?
Janus Core - Our baseline model, trained on a curated dataset of public disclosures across 300+ US financial institutions, is safely and efficiently accessible to all our customers.
Janus Private - The baseline model is then trained on domain- and customer-specific requirements using minimal datasets. The customized models can achieve greater disclosure identification accuracy and recall, further ensuring that the customer remains compliant. These private models are sandboxed and accessible only to a specific customer.
Janus' novel architecture enables the models to tailor to new requirements with an average of just 8–10 additional examples of customer-specific training data. The models also distinguish stylistic and linguistic nuances across financial disclosure text, achieving high accuracy across multiple languages* (no pun intended: this is a disclosure).
🗓 What’s next?
As Janus is deployed into our current disclosure validation pipeline, we are witnessing improvements across the board, enabling us to capture more missing/inaccurate disclosures and provide better recommendations.
We are also working on expanding Janus’ use cases to include providing real-time feedback on agent violations during customer calls and much more!
🙏 Asks!
If you are a financial institution looking for a holistic compliance solution to monitor your marketing and customer interactions, feel free to reach out!
#ai #ml #models #machinelearning #llms #compliance #regulations
* Improvements over baseline approaches, GPT-4, and RoBERTa on our private test dataset.
VP Tech Partners at Nordcloud, an IBM Company
1moMonitoring all metrics, from quality to faithfulness to drift, regardless of the AI platform is a great step forward. Well done IBM watsonx team 👏