Vihayas
Vihayas is a cloud-agnostic infrastructure platform that lets engineering teams define, deploy, and manage cloud resources across AWS, GCP, and Azure using a unified IaC format.
Core trick
Gemini converses to gather requirements, then emits a schema-validated cluster.json and prices it against live cloud pricing for each provider.
Tags
Description
Vihayas is a cloud-agnostic infrastructure platform that lets engineering teams define, deploy, and manage cloud resources across AWS, GCP, and Azure using a unified IaC format.
Today's project called Rachna is the AI design layer built on top of Vihayas. It takes a plain-English description of what you're building — scale, resilience requirements, compliance constraints — and produces a validated Vihayas cluster.json configuration, ready for deployment. Powered by Google Gemini 3.5 Flash. Rachna grounds every decision in real cloud pricing data, generates an architecture diagram, and shows a live cost comparison across all three providers, turning a conversation into production infrastructure.
README excerpt
# Rachna — AI Cloud Infrastructure Design Agent > Built at Google I/O Hackathon 2026 · Vihayas Cloud Platform Rachna is a conversational AI agent that designs production-ready cloud infrastructure from plain-English requirements. Describe what you're building; Rachna asks the right clarifying questions, then generates a validated IaC configuration, an architecture diagram, and a cross-provider cost comparison — all grounded in live pricing data. --- ## Demo  1. **Chat** — Rachna interviews you about scale, resilience, compliance, and budget 2. **cluster.json** — a validated Vihayas IaC configuration appears in the output panel 3. **Diagram** — a Mermaid architecture diagram renders automatically 4. **Compare** — side-by-side cost estimates for AWS, GCP, and Azure --- ## What was built at the hackathon ### `rachna-agent/` — Python backend | File | What it does | |---|---| | `agent/loop.py` | LiteLLM agentic tool-use loop; session management; streaming NDJSON events; loop guards (max 20 iterations, max 3 consecutive identical tool calls); friendly error mapping | | `agent/prompts.py` | Three-phase system prompt: DISCOVERY → GENERATION → REFINEMENT. Guided by a senior principal cloud architect's question framework — derives answers from context, asks only for true information gaps | | `tools/definitions.py` | 5 OpenAI-format tool schemas consumed by LiteLLM | | `tools/handlers.py` | Tool name → async handler dispatch | | `tools/sadhak_