FrootAI — AmpliFAI your AI Ecosystem Get Started
Research-quality data from a production engine. Reproducible. Cite-able. Every number is verifiable — every benchmark ships a runnable script.
Monthly cost delta across 100 enterprise workloads — AVM-composed Bicep vs hand-authored.
Carbon footprint across 200 AVM compositions by Azure region — grid-intensity aware.
Harvest pipeline throughput: Node vs Python vs azd template ingest per language.
Comparison studies with reproducible scripts. Every number is verifiable.
Monthly Azure cost comparison between AVM-composed infrastructure and hand-authored Bicep templates across 100 enterprise workloads. AVM compositions average 12% lower cost due to right-sized SKU selection.
Grid-intensity-aware carbon estimates for 200 AVM-composed templates deployed across 12 Azure regions. Sweden Central and France Central show 60-70% lower emissions than US East due to renewable grid mix.
End-to-end throughput of the 7-stage harvest pipeline across 50 repos. Node implementation processes S1-S7 in 4.2s median; Python in 6.8s; azd template baseline in 12.1s.
Downloadable Parquet datasets — CC0-licensed, schema-documented, weekly-refreshed.
Every harvested Solution Play with provenance fields — source repo, pipeline version, confidence scores, LLM steps, composition metadata. Updated weekly as new plays are harvested.
10 rows · Updated
0 downloads · CC0-1.0
Full Azure Verified Module catalog with classification, WAF pillar coverage flags, cost-band estimates, and popularity ranking. Covers both Bicep and Terraform modules. Weekly-refreshed from upstream registries.
20 rows · Updated
0 downloads · CC0-1.0
Per-play WAF compliance scores across all 5 Well-Architected Framework pillars plus per-CAF-domain scores. Nullable scores where zero applicable checks exist. Weekly-refreshed.
20 rows · Updated
0 downloads · CC0-1.0
Research-quality posts from the engine team. Each tagged shipped / archived / referenced.
The V2 Resolver assigns a confidence score to each candidate AVM module. This experiment documents how we calibrated the scorer against 500 hand-labeled module selections to achieve 92% top-1 accuracy.
View sourceWe ablated the S3 Extract system message — removing one section at a time — to measure which prompt components contribute most to RepoFacts accuracy. domain and services fields drop accuracy by 15-20% when removed.
View sourceCompared three embedding models on the S4 Retrieve stage across 50 repos. text-embedding-3-large achieves NDCG@5 = 0.84, 12% higher than ada-002 (0.75) and 8% higher than E5-large (0.78).
View sourceTested 5 strictness levels of the S7 policy overlay across 100 workloads. 'Strict' mode blocks 18% of deployments but catches 100% of non-compliant SKUs. 'Permissive' mode blocks 2% but misses 35% of violations.
View sourceMapped all 80+ WAF checks against the Azure Well-Architected review criteria. Found 12 criteria with zero automated checks — all in the Cost Optimization pillar. Prioritized for V1.20+ implementation.
View sourcePer §0.5 principle #8: every experiment is tagged shipped, archived, or referenced from production. Quarterly review culls stale entries.
Experiment results integrated into production engine.
Cited by production code or external publications.
Superseded or no longer applicable. Retained for reproducibility.
| Quarter | Before | Culled | Archived | Promoted | Cull Rate |
|---|---|---|---|---|---|
| 2026-Q22026-06-14 | 5 | 0 | 0 | 0 | 0% |
Latest: Initial lab launch — all experiments are new. No culls needed.