Beacon Agent

• 

  screenshot of the working interface

Inspiration

Data platforms don't have a discovery problem anymore — they have a trust problem. Catalogs like DataHub list every dataset, dashboard, model, and pipeline, but the second a new asset lands, someone asks in Slack: "Can I actually use this?" No owner, unchecked PII, half-resolved lineage. We wanted the thing that watches the graph so people don't have to air-traffic control for metadata.

What it does

Beacon; streams new and changed assets in real time and scores each one (0–100) across six governance factors; ownership, lineage, PII sensitivity, documentation, freshness/SLA, and schema stability:

$$ \text{Trust} = \sum_{i=1}^{6} p_i, \qquad \sum m_i = 18+20+22+15+13+12 = 100 $$

An optional autonomous agent you can arm then acts on that judgment computing a separate confidence from trust, lineage $\ell$, and freshness $f$:

$$ \text{conf} = \left\lfloor 100\left(0.5\cdot\tfrac{s}{100} + 0.3\,\ell + 0.2\,f\right)\right\rceil $$

It auto-certifies high-confidence assets, quarantines risky ones, and skips the rest to a human — all gated by a finite action budget so it fails safe. Six screens: Launch Feed, Governance Agent, Watchlist, Alert Rules, Source Health, and a Slack/Telegram alert preview.

How we built it

A design-to-code port. Beacon started as a prototype in Claude's design tool (Beacon.dc.html, Nocturne design system), which we re-implemented as a real Vite + React + TypeScript app — domain logic (scoring, agent policy, generators) in a framework-agnostic module, the Nocturne tokens carried over verbatim, and mock assets streaming client-side every ~2.8s.

Challenges

• Porting a class component's imperative setState into hooks without stale closures inside setInterval — solved with a merge-style setState backed by a hot ref.
• Carrying hundreds of inline style strings into React via a tiny css() parser instead of rewriting each by hand.
• Verifying render (not just compile) with a renderToStaticMarkup smoke test.

What we learned

Separate the judgment from the view, give autonomy a budget so it feels safe, and treat confidence ≠ trust that split is what makes automated governance defensible.

What's next

Wire the scoring and connect/writeback flow to a live DataHub graph over MCP.

Built With

• acryl-datahub
• apache-kafka
• claude
• cloudflare
• css
• datahub
• docker
• docker-compose
• git
• github
• graphql
• html
• javascript
• mysql
• nginx
• node.js
• opensearch
• react
• typescript
• vercel
• vercel-functions
• vite
• vitest