Bridge the gap between AI research and production-grade architecture.
Estimate VRAM, Resource Footprints, and Multi-Cloud OpEx with mathematically-backed precision.
An open-source capacity planning and FinOps estimation tool for production-grade Conversational AI pipelines. This project is built to bridge the gap between AI research and real-world solution architecture, providing a mathematically-backed "compass" for infrastructure design.
We go beyond simple VRAM calculators. We track the entire infrastructure stack:
- GPU VRAM: Weights, KV Cache (vLLM PagedAttention), and component overhead.
- System RAM: Host memory for model runtimes and audio buffering.
- CPU Cores: Dedicated vCPU recommendations based on architecture.
Match your workload with the right hardware using our Required Compute (TFLOPS) engine:
TFLOPS = 2 × Params × Total_Throughput.- Compare your needs against reference points like NVIDIA H100, A100, and L4 to avoid over-provisioning.
Test how your architecture scales across different patterns:
- Monolithic (AIO): Single-instance, all-in-one pods.
- Distributed: Microservices-based pods for STT, LLM, and TTS separately.
Find the most cost-effective cloud GPUs (GCP, AWS, Azure) or calculate the On-Prem ROI (CapEx vs. OpEx) to decide when it's time to build your own cluster.
The FinOps Estimator transforms complex model parameters and concurrency requirements into a professional infrastructure verdict.
graph LR
A[<b>Pipeline Config</b><br/>CCU, RTF, Model Size] --> B[<b>Logic Engine</b><br/>VRAM, RAM, TFLOPS]
B --> C{<b>Cloud Matchmaker</b>}
C -->|GCP| D[G2/A2 GKE Node Pools]
C -->|AWS| E[G5/P4 EKS Node Groups]
C -->|Azure| F[NC/ND AKS Node Pools]
D & E & F --> G[<b>Artifact Export</b><br/>Terraform & HPA Configs]
| Feature | Description | Architect Value |
|---|---|---|
| Multi-Resource Modeling | Tracks GPU VRAM, System RAM, and vCPU Cores simultaneously. | Avoid "CPU Traps" and bottlenecks. |
| Compute Verification | Maps throughput against GPU TFLOPS (H100, A100, L4). | Prevent over-provisioning. |
| FinOps Blueprint | Integrated Spot/Preemptible savings across GCP, AWS, and Azure. | Real-world OpEx predictability. |
| Granular Customization | Independent "Use Custom" toggles for STT, LLM, and TTS. | High-fidelity POC simulation. |
| Infrastructure-as-Code | Dynamic Terraform Export based on architecture verdict. | Rapid baseline provisioning. |
- Mathematically Backed: Formulas derived from vLLM PagedAttention, TensorRT-LLM, and Pipecat benchmark data.
- Consultative UX: Real-time "Solution Architect Verdict" banner that changes based on CCU and precision.
- Production-Ready: Export standard Kubernetes HPA and Terraform configurations directly from the results tab.
- Clone the repo:
git clone https://github.com/JoesSattes/FinOps-Estimator.git - Install:
npm install - Run:
npm run dev - Export:
npm run build(Static export for GitHub Pages)
- Phase 1: Core VRAM & LLM Parameter Logic.
- Phase 2: Multi-Cloud FinOps & Solution Verdicts.
- Phase 3: Granular Customization & Terraform Export.
- Phase 4: Advanced Latency simulation (p95 prediction).
- Phase 5: Multi-Region "Greener-AI" carbon footprint estimation.
This project aims to help everyone understand real AI solution architecture. We welcome contributions from architects and researchers to refine our formulas and data presets.
- CONTRIBUTOR.md: Read our guide to adding new models and pricing.
- Formulas: Based on vLLM, TensorRT-LLM, and Pipecat/LiveKit benchmarks.
March 2026 Update: Hardware pricing and model presets have been synchronized with latest cloud instance data. Join our Contributor Network to help refine the blueprints.