Since early 2021 I’ve been *Digital Senior Product Manager – AI Enablement* in DBT’s Digital, Data & Technology (DDaT) directorate. My small, multi‑disciplinary team looks after two overlapping missions:
1. Ship real products that help colleagues do better work today.
2. Model good practice so other departments can adapt, improve and re‑use.
1 · * Redbox – an "LLM‑in‑a‑box"
In late 2023 a handful of policy advisers asked whether ChatGPT‑style tooling could safely accelerate evidence reviews.DBT collaborated with DSIT’s *i.AI* incubator, mapping user journeys and risk points. The result is **Redbox** – a secure workspace where staff can drop a policy brief, prompt against multiple models, and trace every answer back to source. Redbox@DBT is our custom fork from the core repo, our development is focused on advanced document interrogation and agentic workflows.
Why it matters – It creates more time for policy and analysis colleagues to focus on the deep thinking rather than the mundane, whilst giving senior leaders confidence that nothing sensitive is leaking to the public internet.
How it’s built – GOV.UK Design System front‑end on Django, Anthropic and Amazon models via the secure AWS Bedrock platform on DBT infrastructure, Langchain as the core for agentic design.
Where we are – Live across DBT; sister installs now running in Cabinet Office and DSIT. We run open show‑and‑tell sessions every fortnight.
2 · Leaving DBT Service – automated off‑boarding across the organisation
Off‑boarding used to mean emails, spreadsheets and nervous rummaging for lost kit. We replaced that with a single, guided journey: line managers answers a few questions, the system handles the rest. Accounts close, laptops get booked for courier pick‑up, pay stops on the correct day, and leavers have space to write a knowledge‑handover note.
The result? Less last‑minute panic and a cleaner, real‑time picture of our workforce.
3 · Building the guard‑rails
Alongside product work I contributed to DBT’s **Generative‑AI Governance Framework** – a lightweight checklist covering data‑classification, model selection, prompt hygiene and human sign‑off. We share it under an open licence because good guard‑rails should be easy to copy.
Analysis‑in‑Government Month 2025 – "Redbox: Generative AI in Government".
Guest posts on the DBT digital blog, LinkedIn essays on "LLM UX" and working in the open.
---
Pragmatic first – tiny, well‑scoped pilots beat grand visions.
Open by default – code, docs and decks are published unless there’s a good reason not to.
People in the loop – AI augments, never replaces, human judgment.
*If you’re building something similar or want to compare notes, feel free to get in touch.