AI Decision Layer
Every deck distils the live EXAFABS snapshot into four concrete recommendations — where to deploy capital, what's fragile in the supply base, who is climbing fastest, and which competitive lines are closing. One card per kind, regenerated on demand.
Today's Deck
Refreshed from the current Convex snapshot. Cached server-side for up to 5 minutes across all analysts.
How the deck is generated
1. Snapshot
We read the live fabs_plants, fabs_stats, and fabs_insights collections from Convex and build a focused aggregate: top countries by advanced capacity, top companies by share, node mix, and recent intelligence.
2. Reasoning
Claude Haiku 4.5 receives the aggregate and a tight JSON schema. It must reference concrete numbers from the context for every bullet — answers that cite no data are rejected by the server parser.
3. Validation
The server enforces exactly four kinds, clamps recommendation length, trims rationale bullets, and validates confidence grades. Missing kinds are filled with placeholder cards instead of collapsing the grid.
4. Caching
Decks are cached server-side for 5 minutes so ten analysts refreshing at 9am see the same deck without 10× the cost. Rate limited to 3 fresh generations per IP per hour.
Card kinds
Investment Signal
Where capital should flow next
Supply Risk
Concentration & single points of failure
Growth Opportunity
Fastest-rising regions and players
Competitive Shift
Who is closing the gap