Resources

Educational content

Short explainers on the concepts behind our work, in plain language.

E/01

What is a prediction transformer?

How a model like PE-1 reads time the way a language model reads text.

Key points

  • Time series are sliced into patches and treated as tokens
  • A decoder-only transformer predicts the next stretch of the future
  • Pretraining on billions of observations enables zero-shot forecasts
  • Outputs are distributions with calibrated uncertainty
E/02

RAG vs fine-tuning

Two ways to make a model know your business, and when to use each.

Key points

  • RAG retrieves your documents at question time: fresh, auditable, cheap to update
  • Fine-tuning changes the model's behaviour: tone, format, specialized tasks
  • Most assistants start with RAG and add fine-tuning later
  • Both need clean data more than clever prompts
E/03

Measuring automation ROI

A simple formula that survives contact with finance.

Key points

  • Hours saved × loaded hourly cost = gross saving
  • Subtract build, licence and maintenance costs
  • Add error reduction and speed gains when you can measure them
  • Decide on a 12-month horizon and measure monthly
E/04

Reading a forecast's confidence band

Why a single number is a trap and how to use P10–P90.

Key points

  • P50 is the central estimate, not a promise
  • P10–P90 covers the range you should plan for
  • Wide bands mean uncertain data, not a broken model
  • Decide with the band: safety stock, staffing, budget
E/05

Data quality basics

The four checks that fix most analytics problems before any AI is involved.

Key points

  • Completeness: no silent gaps in key fields
  • Consistency: one definition per metric across systems
  • Timeliness: data arrives when decisions are made
  • Lineage: you can trace every number to its source
E/06

Glossary

The vocabulary you will meet in our proposals.

Key points

  • LLM: a large language model that reads and writes text
  • Agent: a model that takes actions through tools, not just answers
  • Embedding: a numeric representation used for semantic search
  • Zero-shot: forecasting or classifying without task-specific training
  • Sovereign cloud: hosting under local jurisdiction and control