Shipping Production AI Inside Government — William Tarr, Ministry of Justice (DO NOT PUBLISH)

William Tarr and Louis Hogarth from the UK Ministry of Justice (MOJ) Justice AI Unit present a frank account of what it actually takes to ship production AI inside a 95,000-person government department. Tarr, a forward-deployed engineer who dropped out of Harvard and ran a YC-backed company before joining government via the Number 10 Innovation Fellowship, describes the unit as "the closest thing you can find to a startup in government." Founded roughly a year and a half before the talk, the unit has grown to approximately 40 people — a headcount that Tarr argues produces the output of a conventional 300-person government team.

The talk opens with the context that motivated the unit's creation: daily headlines about prisoners being wrongly released from Wandsworth Prison, a symptom of disjointed systems across courts, probation, and prisons, where different departments use different identifiers for the same individual, legacy tech stacks are connected by fragile APIs, and frontline staff and HQ policy makers have almost no shared understanding of actual workflows.

The unit's defining methodology is the forward-deployed model. Both engineers spend two to three days per week physically on site — inside prisons with keys, in probation offices — rather than working from policy documents or roadmaps. Tarr notes that "policy documents that outline best use and perfect operation are just tacitly followed in a prison" and that real workflows bear little resemblance to the idealized versions described at HQ. Hogarth describes a probation officer tool built by just two engineers that went from MVP to national rollout in a matter of months — a cycle described as "unheard of in government."

Building in this environment surfaces infrastructure constraints that nobody at headquarters knows about: many Victorian prison buildings have six- to seven-foot walls that block WiFi entirely, and in some facilities WiFi is prohibited because prisoners could access it. The team's response was to ship offline mode via browser-level caching, rather than waiting on infrastructure changes. Similarly, when they encounter connectivity gaps during field work, they pivot to on-premises or fully offline deployments on the same visit.

On governance, the talk draws a sharp distinction between governance designed for policy (long committee processes, months of deliberation by people who have never been inside a prison) versus governance fit for products. The team's strategy for navigating this is to avoid theoretical pitches entirely: they run safe-to-fail pilots, collect real user signal, and arrive at governance meetings with observable behavioral change rather than slides. Feature flags are used to test at small scale before broader rollout; code is killed without hesitation when something does not land.

The roadmap ahead includes CCTV analytics and predictive intelligence (moving from reactive to proactive use of prison camera footage), AI-assisted report writing and case note automation so probation officers spend more time with clients and less time on paperwork, transcription of spoken interactions, and broader data integration to eliminate the kind of release-error failures that motivated the unit's formation. Education and AI literacy for frontline staff is also called out — staff are actively engaged on site with short videos, plain-language communication, and in-person walkthroughs rather than mandatory AI seminars.

The talk closes with a direct hiring pitch: the Justice AI Unit is recruiting four additional forward-deployed engineers to work embedded across the MOJ. The core message throughout is that proximity to real users — not engineering sophistication alone — is what makes fast, impactful government AI possible.