KALKI AI — Business Model, Cost Estimation & Scalability Plan
1. Cloud Infrastructure Cost Estimation (Monthly Projections)
| Component |
Resource Specification |
Unit Cost |
Projected Monthly Cost (100k Users) |
| GPU Inference Cluster |
8x NVIDIA H100 SXM (vLLM Engine) |
$3.50 / GPU-hr |
$20,160 |
| Vector Storage (Qdrant) |
3-Node Cluster (128GB RAM, SSD) |
Managed Cloud |
$1,200 |
| Relational Database |
AWS Aurora PostgreSQL (Multi-AZ) |
db.r6g.2xlarge |
$1,450 |
| Caching & Pub/Sub |
Redis Enterprise Cluster |
32 GB RAM |
$450 |
| Object Storage |
AWS S3 (Document storage & backups) |
10 TB |
$230 |
| Total Estimated Infrastructure Cost |
|
|
~$23,490 / month |
2. Horizontal Scalability Roadmap
graph LR
Tier1["Phase 1: Single Cluster<br/>10k DAU<br/>Monolithic Gateway + Qdrant"] --> Tier2["Phase 2: Microservices<br/>500k DAU<br/>Ray LLM Cluster + Distributed Redis"]
Tier2 --> Tier3["Phase 3: Multi-Region Global Mesh<br/>10M+ DAU<br/>Anycast Edge SLMs + Cloud MoE Fallback"]
3. Risk Assessment & Mitigation Matrix
| Risk Factor |
Severity |
Probability |
Mitigation Strategy |
| Prompt Injection Attack |
High |
Medium |
Dual-pass Security Agent verification using Llama-Guard 3 + static input sanitization. |
| Hallucination in RAG Outputs |
High |
Medium |
Validator Agent verification threshold ($>0.85$ grounded confidence rating) before sending output. |
| GPU Capacity Exhaustion |
Medium |
High |
Automatic graceful degradation to quantized local Edge SLMs or speculative decoding. |
| Privacy Data Leakage |
High |
Low |
Dynamic PII anonymization regex layer on all document ingestion pipelines. |