OmniArch: The Cloud Agent

Inspiration

As a Senior Google Cloud Architect working closely with government and enterprise clients, I realized a painful truth: only a fraction of my time is actually spent designing cloud systems. The rest is heavily consumed by administrative friction and "translation" tasks.

I constantly found myself:

• Translating complex technical constraints (like why Google's OIDC doesn't support prompt=login) to non-technical security executives.
• Reverse-engineering budget requirements into valid architecture sizings (Volumetrics) to pass strict vendor funding audits, such as Google's Partner Service Funds (PSF).
• Navigating vendor support labyrinths, repeatedly explaining the difference between a Google Sales Payment ID and a GCP Console Billing ID.

I wanted to build a "second brain"—an AI that bridges the gap between deep technical truth and business delivery, freeing architects to do what they do best: innovate. This inspired the creation of OmniArch: The Cloud Agent.

What it does

OmniArch serves as a multi-functional Copilot specifically designed for Cloud Partners and Senior Architects. It acts across three core domains:

• Architecture & Security Translator: It instantly translates complex protocol limitations into actionable, secure design alternatives. For instance, when a client demands forced MFA re-authentication via Google SSO, OmniArch generates a professional pitch explaining the limitations of OAuth 2.0 and proposes App-Level Step-Up Authentication workflows.
• Proposal & SOW Automation: It analyzes technical requirements and automatically refines Statement of Work (SOW) documents. It knows vendor "unwritten rules" (e.g., converting "Effort/Days" into "Deliverables" and "Fair Market Value" for PSF funding) and can mathematically deduce realistic infrastructure volumetrics (GKE node sizing, Cloud SQL specs) based on target Annual Recurring Revenue (ARR) and Non-Functional Requirements.
• Vendor Operations Manager: It drafts precise, context-aware support cases to resolve backend vendor issues without the usual back-and-forth friction, knowing exactly what screenshots or IDs to provide.

How we built it

We leveraged the power of OpenAI's advanced models to build a highly specialized agent:

• Core Engine: Powered by the latest OpenAI models for superior reasoning and natural language processing capabilities.
• Retrieval-Augmented Generation (RAG): We compiled a proprietary Knowledge Base consisting of real-world SOW templates, past vendor rejection feedback, and deep-dive FAQs on cloud security protocols. We formatted this data into structured Markdown/JSON and fed it into the OpenAI Assistants API utilizing File Search capabilities.
• Prompt Engineering: We crafted strict System Instructions defining a "Senior Cloud Architect" persona. This ensures the Agent doesn't just regurgitate generic public cloud documentation, but provides partner-specific consultancy.

Challenges we ran into

• Balancing Technical Depth with Business Language: It was challenging to tune the AI to speak technically enough for engineers, yet diplomatically enough for C-level executives. We solved this through iterative prompt tuning and contextual few-shot examples.
• Hallucination in Architecture Sizing: Initially, the AI would guess random server sizes. We overcame this by injecting strict reasoning steps into the prompt, forcing the model to calculate the baseline load mathematically. For example, to estimate the fair market value cost $C$ for a task, we strictly guide it to compute $C = H \times R$, where $H$ is the total hours and $R$ is the hourly rate. To estimate infrastructure nodes $N$ required based on user traffic $U$, we apply a baseline equation such as $N = \lceil \frac{U}{C_{node}} \rceil$.

Accomplishments that we're proud of

We successfully transformed a collection of disjointed, real-world frustrations into a cohesive AI tool. OmniArch doesn't just write text; it understands the business context of being a Cloud Partner. Getting the Agent to successfully transform an SOW "Effort" table into a compliant "Fair Market Value" appendix was a massive win for productivity.

What we learned

• The true power of RAG lies not just in the volume of data, but in the structure of the knowledge base. Cleaning up our past SOWs and FAQs into a structured format drastically improved the Agent's accuracy.
• System Prompting is an architecture design process in itself. Defining boundaries (what the Agent shouldn't do) is just as important as defining its capabilities.

What's next for OmniArch: The Cloud Agent

• Google Workspace Integration: Integrating the Assistants API directly into Google Docs and Gmail, allowing OmniArch to review and comment on SOWs in real-time within the document itself.
• Multi-Cloud Expansion: Expanding the knowledge base to include AWS and Azure partner funding frameworks and architectural best practices.
• Automated Architecture Diagramming: Integrating with tools like PlantUML or Draw.io to generate visual architecture diagrams directly from the text-based SOW requirements.

Built With

• gcp
• python