From job-risk dashboard to research engine.
Skynet Jobs started as a static public tracker for AI labor-market risk. V2 turns it into an auditable research pipeline while keeping the production site fast, static, and easy to trust.
Project brief
A public command center for tracking how AI changes work.
The site helps visitors search occupations, compare automation exposure, inspect productivity signals, and follow evidence that the labor market is shifting. The V2 work is about credibility: dated snapshots, source trails, and a repeatable research loop instead of one-off dashboard edits.
V1 - current foundation
Static tracker
Public experience
A fast static React dashboard for searching occupations, reading risk scores, and scanning labor-market signals.
Data model
Curated TypeScript data files and generated site metrics committed with the frontend. The browser reads a static bundle, not a live database.
Deployment
A source-control workflow builds the app, publishes static assets to object storage, and refreshes the CDN cache.
Newsletter path
A small serverless API records opt-in signups in a managed key-value store and keeps secrets out of the client bundle.
V2 - in progress
Research ledger
Nightly research agent
A local GPU workstation runs a scheduled agent that collects sources, asks for structured analysis, and produces candidate updates.
Evidence ledger
Managed NoSQL tables store role snapshots, evidence links, layoff events, research runs, and dated predictions for auditability.
Review gate
The agent can propose changes, but site-facing data is approved before it becomes part of the public static export.
Static production export
Approved records are transformed back into generated frontend data so the live site stays fast, cacheable, and resilient.
Data pipeline
The live site consumes reviewed snapshots, not raw agent output.
The important design choice is separation. The research system can be messy, iterative, and evidence-heavy. The production site receives a clean static export after review, so visitors get a stable experience and every number can still be traced backward.
| Ledger area | What it proves |
|---|---|
| Role snapshots | Risk score, sector, trend, confidence, and model rationale by date. |
| Evidence records | Source metadata, citation notes, extraction status, and reviewer disposition. |
| Layoff events | Company, date, affected count, AI relevance classification, and evidence links. |
| Predictions | Forecast, horizon, confidence, source basis, and later outcome tracking. |
| Research runs | Agent version, prompts, model choice, run status, and review summary. |
V2 flow
Research, review, export, deploy.
01
Collect
Nightly agent gathers source material and extracts candidate signals.
02
Analyze
Local model or external API produces structured notes with confidence and uncertainty.
03
Review
Human approval decides which changes become public facts or predictions.
04
Publish
Generated static data ships with the frontend through the existing deployment path.