AI and data

We agree an evaluation set with your team before building: real examples with the answers your experts expect. Every version of the system is scored against it, and reviewers check a sample of the harder cases. You decide the threshold for release. In production, we keep measuring and route uncertain cases to a person rather than guessing.

Read more: AI developmentLink to this question

We work with commercial models such as OpenAI GPT models, Anthropic Claude and Google Gemini, with open models such as Meta Llama and Mistral, and with classic machine learning where it fits better. We choose per task on quality, cost, latency and where the model may run. We design so the model can be changed later. Naming a vendor does not mean a partnership.

Read more: AI developmentLink to this question

We reduce the risk in layers. The model answers from retrieved passages of your content and must cite them. Guardrails check that claims are supported by those passages. When the sources do not contain an answer, the feature says so instead of guessing. The evaluation set measures how often answers are unsupported, and we do not release until that rate meets the threshold you agree.

Read more: Generative AI and LLM integrationLink to this question

For answering from your documents, retrieval is usually the better start: it keeps answers current, shows sources and respects permissions. Fine-tuning helps when you need a consistent format or style, a specialised vocabulary, or a smaller model that runs cheaply. Many systems use both. We test the options against your evaluation set and recommend the simplest one that meets the bar.

Read more: Generative AI and LLM integrationLink to this question

We choose per task, based on answer quality on your evaluation set, cost per request, speed, language support and where the model is allowed to run. Options include OpenAI GPT models, Anthropic Claude, Google Gemini, Meta Llama and Mistral, through services such as Azure OpenAI, Amazon Bedrock or Vertex AI. The model sits behind one interface, so it can be replaced later.

Read more: Generative AI and LLM integrationLink to this question

It depends on the problem, not on a fixed amount. What matters is enough history to cover the patterns the model must learn, including seasons and rare events, and reliable labels for the outcome you want to predict. We check this during the data review and tell you plainly if the data cannot support the target yet, and what to collect so it can.

Read more: Machine learningLink to this question

It depends on your cloud, volumes, skills and budget more than on any single product. Snowflake, Databricks, BigQuery and Amazon Redshift all work well for most companies, and PostgreSQL is enough for many. We compare the options against your requirements and existing contracts, and design so the transformations, written in SQL with dbt, stay portable.

Read more: Data engineering and analyticsLink to this question

Yes. We use the ERP’s supported extraction methods, such as its APIs, export services or change data capture from a replica, rather than heavy queries against the live database. Loads are scheduled outside peak hours where needed. For SAP, Oracle, Microsoft Dynamics 365, NetSuite and Odoo we follow each vendor’s recommended approach.

Read more: Data engineering and analyticsLink to this question

Only for decisions that cannot wait, such as fraud screening, stock availability or payment status. For most reporting, loads every hour or overnight are enough and cost far less to run. We use streaming with Apache Kafka or change data capture where real time is needed, and batch where it is not, and we decide per use case with you.

Read more: Data engineering and analyticsLink to this question

Most structured and semi-structured business documents: supplier invoices, credit notes, receipts, bank statements and advices, purchase orders, delivery notes, customs and shipping documents, claims, application forms and identity documents. Handwriting, poor scans and unusual layouts lower accuracy, which is why we measure on your own samples before agreeing the scope.

Read more: Intelligent automation and document AILink to this question

No. Robotic process automation copies clicks on a screen and breaks when the screen changes. We read the document with document AI, check it with business rules and post through the ERP’s supported interfaces. Where a system has no interface at all, screen automation can be a last resort, and we say so when we use it.

Read more: Intelligent automation and document AILink to this question

We post to SAP S/4HANA, Oracle NetSuite, Microsoft Dynamics 365, Odoo and Sage, and to custom and workflow systems, through their APIs, import services or integration platforms. Posting follows the same approval rules your team uses today. Naming a product here does not mean a partnership with its vendor.

Read more: Intelligent automation and document AILink to this question