The Decision Engine
Six stages from an English sentence to a defensible infrastructure plan, with its reasoning left visible.
GPUVerse’s core differentiator: given a workload and its constraints, it returns ranked infrastructure configurations with explanations, deterministic, weighted, and honest about its confidence.
Shipped
A deterministic constraint engine, an evaluation engine, and a stable ranker produce explainable, immutable RecommendationReport objects (shipped 2026-07-19).How a recommendation is made
A workload enters at the top; a plan resolves at the bottom. The whole path is deterministic, the same inputs always produce the same decision, and every stage is inspectable.
Describe workload
Model, size, scale, latency, budget, compliance
Understand requirements
GPU memory, parallelism, intensity
Analyze constraints
Budget, region, compliance, GPU count
Evaluate providers
Score every candidate on five metrics
Compare GPUs
Rank against your objective
Estimate cost
Base, egress, spot-interruption-adjusted
Generate infrastructure plan
The explained decision
How it scores
After filtering candidates by hard, capability, and business constraints and sizing them by VRAM, the engine scores five metrics, cost, performance, reliability, availability, and compliance, into a composite score. Default weights (configurable per organization):
| Dimension | Weight | What it measures |
|---|---|---|
| Performance fit | 40% | How well GPU specs match the workload |
| Cost efficiency | 30% | Price-to-performance ratio |
| Availability | 20% | Spot interruption rate, current stock |
| Compliance | 10% | Certifications match requirements |
The Decision Engine layer
A Decision Engine (ADR-013) sits over the recommendation engine. It reuses the engine and never re-ranks, the deterministic pick stands, but adds the two product-promised dimensions the engine lacked: scalability and operational complexity. Together with cost, performance, availability, and compliance, that makes six scored dimensions, each with a score, weight, explanation, and factors.
Explainability & confidence
Every recommendation includes evidence, tradeoffs, risks, and a selection narrative, plus 2–3 concise bullet points on why a configuration was chosen. Confidence is provenance-aware, a confident recommendation built on an unverified price is downgraded, because it would otherwise be dishonest.
0.9+
High confidence
Clear best option, strong data
0.7–0.9
Medium confidence
Solid, with tradeoffs
< 0.5
Experimental
Flagged, sparse or volatile data
Performance targets
Recommendation latency P50 < 200ms, P99 < 1s; accepted rate > 60%; explanation quality > 4.0/5.0.Was this page helpful?