UK AI Radar: An Agentic Radar Over Government Surface Area

#ai#agents#playground+2 more

Illustration — UK AI Radar: An Agentic Radar Over Government Surface Area

Repository: hyperdrift-io/uk-ai-radar · Live: radar.hyperdrift.io · Status: POC, open in public · Not affiliated with HM Government.

The missing layer

UK AI founders have to track an extraordinary amount of public surface area to do their jobs: department announcements on gov.uk, open grants on UKRI and the Innovate UK Funding Service, Parliamentary committees and Hansard, consultations, policy papers. Each of these systems is exhaustive in its own way. None of them is personalised. The newsletters that grew up around the gap go the other way: generic, and stripped of citation.

The missing layer sits in between: a personalised, cited, multi-source radar that knows your profile and tells you what changed this week that you should act on.

UK AI Radar is a POC for that layer. It is also a worked example of a pattern that generalises: a small graph of typed personas reading authoritative sources through a diff-first cache, emitting a brief ranked against an explicit founder profile. Every claim links to a source the reader can check.

What it does, concretely

You drop a JSON profile under profiles/:

{
  "name": "Acme Robotics",
  "stage": "seed",
  "focus": ["robotics", "manufacturing", "industrial-ai"],
  "regions": ["uk", "eu"],
  "interests": [
    "innovate-uk grants relevant to robotics",
    "skills and immigration policy affecting AI hires",
    "procurement frameworks where we could be a supplier"
  ]
}

You run npm run digest -- --profile profiles/acme.json, and you get back a weekly brief ranked by relevance to your profile. Each item says what changed, why it matters to you, and links to the gov.uk, UKRI or Parliament page that says so.

A second profile on the same cache produces a visibly different top three. Same surface area, different lens.

The architecture, and why it generalises

The internal shape is a small typed agent graph:

PersonaJobInputOutput
ScoutFetch, hash, diff. Forward only deltas downstream.Whitelisted source URLsRaw page deltas + content hash
AnalystExtract structured fields against Zod schemas (Anthropic tool-use).Page deltaTyped records (grants, consultations, statements, …)
StrategistScore each record against the founder profile. Write the "so what?" angle.Typed records + profileRanked, annotated items
EditorDedupe, rank globally, emit the brief.All Strategist outputsOne Markdown brief, every claim cited

Three properties of this shape recur in every agentic system I've built.

1. Personas are not chat roles. Each persona is a typed node in the graph with a tight input and output schema. Calling it Scout rather than fetch_and_diff() isn't anthropomorphism. It makes the role one boundary, with one responsibility, that you can swap without touching the others. A new engineer can read the system from the graph.

2. The cache is the difference between toy and product. The Scout hashes each page, stores the hash in SQLite and forwards only deltas downstream. On the first run every page is new and the token bill is significant. On a second run an hour later, most pages haven't changed, so almost nothing reaches the Analyst. The acceptance bar is a second run under a fifth of the first run's tokens. That is what makes a weekly run per tenant affordable.

3. The whitelist is enforced at the fetcher, not by the agent. Any host outside gov.uk, ukri.org or parliament.uk is rejected before a request goes out. The agent couldn't exfiltrate to an arbitrary URL even if a prompt injection tried to make it. Sources are a property of the system, not of the model: the model only ever sees content that was already inside the trust boundary.

Multi-tenancy on a single laptop

Profiles double as tenants. The same cache and the same graph produce different briefs for different profiles:

npm run digest -- --profile profiles/acme.json
npm run digest -- --profile profiles/beta.json

The Scout fetches once. The Strategist scores twice, and the Editor emits two distinct briefs. This is the TenantContext pattern at its smallest: a profile id flows through every node, scoring uses it, and output is keyed by it. Move that profile from a JSON file to a row in a tenants table backed by your billing system and you have a multi-tenant SaaS.

What it deliberately leaves out

It isn't a government service. There's no crown, no GOV.UK wordmark and no gov.uk subdomain. It uses the GOV.UK Design System because it is excellent and accessible, and a banner on every page makes the non-affiliation explicit.

It doesn't replace the sources. Every claim links to the authoritative page, so the brief helps you read the government's own output rather than standing in for it.

There's no prompt box either. A radar that puts a chat box at the top of the brief is two products glued together, neither of them good. Since September the live site goes the other way: open radar.hyperdrift.io/explore with an agent in your browser and the page offers it its own tools over WebMCP, while you keep the final call on each item. The Agent Is the Session tells that part.

The wider gap

UK AI founders are a small part of a much larger problem: knowledge workers in any regulated, fast-moving field lack personalised, cited radars over their authoritative sources. Healthcare professionals tracking MHRA and NICE. Lawyers tracking case law and statutory instruments. Financial-services teams tracking PRA and FCA bulletins. Defence contractors tracking procurement notices. Researchers tracking funder calls across UKRI, Horizon and NIH.

Each of these audiences has a high-volume, authoritative, public source ecosystem, a need for personalisation, a hard requirement for traceable citation, and a regulatory or competitive cost to missing something that mattered. Most make do with generic newsletters, hand-curated Slack channels and brute-force RSS.

UK AI Radar is a single instance of a pattern that should exist a dozen times.

What's open for adopters

The repo is public: hyperdrift-io/uk-ai-radar. It runs on Nuxt 3 with Vue and TypeScript, LangGraph.js with the Anthropic SDK, SQLite for the page-hash cache and Vitest for tests. A few things are deliberately small:

  • One profile schema, easy to extend.
  • Four personas in v1. Lawyer, Lobbyist, Grant Writer and Archivist are designed in but deferred to v2.
  • One brief format. The architecture allows other renderings, such as an audio brief or an RSS feed, without touching the graph.

The acceptance bar for the POC is recorded so adopters can check the same shape works for them:

  • 100% of cited URLs in a brief resolve to gov.uk, ukri.org or parliament.uk.
  • A second consecutive run consumes < 20% of first-run tokens.
  • Two profiles on the same cache produce visibly different top threes.

If you're building a radar for a different field, fork it: replace the whitelist and the persona prompts, and keep the shape.

Why I wrote this in public

The first reason is the open-source tradition: a pattern other builders can adopt outlives any individual radar.

The second is civic. The UK government's surface area is a public good, paid for by everyone, and legible only to those with time to read it. A radar that brings it within reach of founders who couldn't otherwise keep up flattens, in a small way, an information gradient that shouldn't exist. That's why this project sits in the Playground rather than behind a service tier.

If you are working on something adjacent, or want to discuss adapting the pattern to your field, get in touch.

Let’s talk

What could we build together?

Your note goes to the Hyperdrift team. We’ll use your email to reply about this enquiry.