Ben Grime

Work

Things I've built, and what they did

Each case study says what the problem was, what I built, the decision that mattered most, and what I'd do next. Where something is still in progress, it says so.

  1. 01 Stable Future · 2026Open source Opportunity scanner Where could AI save time in your organisation? Pick an industry, adjust the roles, and see where staff time goes. An agent maps your own job titles to UK occupations and stops to ask when it isn’t sure.
    • LangGraph
    • MCP
    • evals
    • AWS Lambda
    • Terraform
  2. 02 Stable Future · 2026 AI exposure for every UK job US research scores AI exposure task by task, but UK jobs are coded differently. I built the bridge between the two and a score for each of 412 UK job groups.
    • Python
    • LLM agents
    • crosswalks
    • pytest
  3. 03 Waterman Group · 2025–26 Environmental summaries, drafted in Teams Consultants drop a chapter into a Teams chat. The bot finds similar past chapters and drafts the plain-English summary in their style. Ten minutes instead of two hours.
    • RAG
    • Azure OpenAI
    • Bot Service
    • TF-IDF
  4. 04 Research note · September 2026 Are AI-exposed jobs losing young UK workers? A UK test of Stanford’s “canaries in the coal mine” result. The first chart said yes. The checks said not yet.
    • ONS survey data
    • regression
    • bootstrap
  5. 05 Waterman Group · 2025–26 The first AI tools at an engineering firm An inspection app for 300 staff, training that got two non-technical teams building with AI, and a board demo of what agents could already do.
    • Power Platform
    • training
    • agents
    • board demo
  6. 06 BlueDot Impact grant · in progressIn progress GPTs are GPTs, re-rated for 2026 The 2023 study that labelled every US job task for AI exposure, run again on all 18,838 tasks in O*NET 31.0 with three of today’s models and an updated rubric.
    • LLM labelling
    • inter-rater agreement
    • OpenRouter

Earlier

theMakings · 2024–25

Data scientist

A leadership coaching firm with data on 28 clients, 115 cohorts and 5,300 sessions. I used machine learning to cut a 33-question survey to 3 headline measures and built a self-serve Power BI dashboard. Cohort reports went from 3 hours to 2 minutes.

University of Oxford · 2024

AI/ML researcher

Made a synthetic copy of a 3,428-patient malaria trial so researchers could share it without exposing patients. A gradient-boosted classifier could not tell it from the real data.

NHS · 2023–24

Statistician

Analysed patient data for a senior surgeon: a bigger gap between the torn ends of a ruptured Achilles does not rule out treatment without surgery, especially in younger patients. Co-authored the paper in the Journal of Clinical Orthopaedics and Trauma.