Working with AWS, Azure or Google Cloud? No problem. We work with all large language models and AI service providers.

MSBC AI is a container on your architecture under your control, connected to your tech stack and fitted with the latest security controls.

Great question. Licensing costs can be expensive and running costs may be hard to budget. We use open source models pre-trained for specific tasks and with no licence fees.

We do work however, with AWS, Azure, Google Cloud and all mainstream providers.

We use Retrieval Augmented Generation (RAG) to add your data to improve responses without retraining the model.

We also fine-tune by retraining models on new data. Unnecessary elements may be removed to speed response times and increase accuracy.  You cannot retrain close source models such as ChatGPT and Gemini.

In summary, open source models with access to your internal data will cost less to run and perform faster than large models trained to do everything.

Models are only as good as the data. McKinsey says 70% of GenAI adopters struggle with data quality. Preparation may be as much as 80% of the work in automating business processes. MSBC has been doing this for 20 years.

Our client engagement begins with an analysis of current processes. These are compared with the preferred outcomes. We then design a plan to bridge the gap.

Workflows may be as simple as numbers entered into a spreadsheet. This data has a source, even if it’s a telephone call from the factory. We build systems to streamline this work while collecting data. This results in consistent, quality data, free from human error.

Internal data may be combined with third-party sources. This may be market data for investors, weather patterns for insurers, or traffic updates for logistics firms. If new data is required, we are often able to source this from the internet.

The technical skill is combining these different data sources into a single data lake. Once combined, the data is used to train, test and run AI models. Assistants allow everyone to interact with the data, using straightforward questions and answers.

By limiting the data used in each model we control costs and speed up responses. The risk of random and wrong answers is reduced by training models to do specific tasks. These are specific to our clients’ businesses and designed to deliver optimal outcomes.

The data, the models, the assistants and their output may all be housed in a secure environment on clients’ systems. Once the work is completed, all of these are owned by our clients.

Hallucinations are factual inaccuracies or non sensical data. Intrinsic issues occur when the model extracts inaccurate information from source data. Extrinsic issues involve introducing new data.

Flawed training data is the core problem. If models are overfitted to the training data they are less capable of handling new sources. The models may not understand the context in which the data is used and the sampling techniques may cause errors.

The best way to address poor data is to use in-house sources. We use Retrieval Augmented Generation (RAG) and fine-tuning to increase response times and accuracy as well as lower running costs.

In summary, designing a custom model for a specific workflow and training on internal data will eliminate most errors. The model will also cost less to run and perform faster than large models trained to do anything.

Close source models such as ChatGPT and Gemini have access to the internet and may retrieve data where authors and creators claim copyright. Legal cases along these lines are setting precedents and the situation is fluid.

Custom models using internal data are unlikely to face intellectual property issues. They are also cheaper and faster to run than larger models with those vulnerabilities.

The intellectual property in your AI system is 100% yours once it is fully developed and paid for.