AI agents that hold up in production
Most AI projects stall between demo and production. We build agents on Gemini and Claude, test them against your own cases before they reach users, and put cost and safety controls in place from the start.
What we do
- Agentic engineering: multi-agent systems built with Google’s Agent Development Kit (ADK) and Model Context Protocol (MCP) on Cloud Run.
- Apigee AI Gateway: govern LLM traffic with Model Armor and token quotas, and expose your APIs and BigQuery data to agents as MCP tools.
- Gemini Enterprise for CX: shopping assistants and contact centre automation grounded in live BigQuery data.
- Operational AI agents: agents for DevOps, security triage and document processing.
- Agent reliability engineering: continuous evaluation, hallucination detection and circuit breakers on tool calls.
- Claude on Google Cloud: Claude running inside your Vertex AI perimeter.
- Chatbots and search: answers grounded in your Confluence, Jira and Slack content.
From our own work
Ending alert fatigue with an agent fleet: how we built AI investigation agents on Google Cloud that triage production alerts in under 30 seconds and cut engineer paging by 80%.
Key Deliverables
AI & Machine Learning case studies
Named clients and the results they agreed to publish.
Cross-Cloud AWS to Google Cloud Modernization for Fitness Platform Hapana
Migrated millions of active workouts, global member billing, and IoT facility door access controllers from AWS to a secure Google Cloud Run landing zone with zero cutover downtime.
1,000-Core On-Demand Supercomputer for Global Engineering Leader
Engineered an elastic Slurm HPC cluster on Google Cloud with Scale-to-Zero automation, delivering 10x faster simulation turnaround with zero idle compute waste.
Confirm Control: Real-Time Field Risk Governance & Compliance
Architected an offline-first Flutter mobile application with serverless Google Cloud Firestore backend to digitize field hazard logging and APRA-compliant risk governance.
AI & Machine Learning: questions we get asked
What counts as production AI rather than a pilot?
Something with monitoring, evaluation, cost controls, an on-call owner, and a security and governance posture your risk team has actually signed off. Most AI work that stalls stalls because it was built as a demo, and nobody could say what happened when it gave a wrong answer at 2am or who approved it holding that data.
Do you build on Gemini, Claude, or both?
Both, and open models too. Agents run on Google’s Gemini Enterprise Agent Platform, and Claude runs inside your own Google Cloud tenant through Vertex AI, so prompts and context never leave your boundary. Open-weight models run there as well. Staying on Google is what lets us pick the best model for each job and change it later without re-platforming.
How do you stop an AI agent going wrong in production?
Evaluation before release and reliability engineering after it. We treat agents as services: defined inputs, measured outputs, alerting on drift, and a rollback path.
Can you work inside our existing data governance?
Yes. Data residency, access controls and audit trails are part of the build, not a retrofit. For APRA-regulated clients we deliver against CPS 234, and we use Wiz to evidence the security posture rather than asserting it.
What does an AI engagement cost?
An agent MVP in your own environment is $25,000 and takes two weeks. Moving that agent to production is quoted after the MVP, because the cost depends on what it touches.
How quickly can we get something in front of users?
Two weeks for an agent MVP in your environment. Whole-platform work is longer and we will say so before you sign anything.
Talk to an architect who has done this before.
Bring your current setup and the outcome you need. You will get a view on the approach, the risks and roughly what it costs.
Straight to a senior GCP architect. No SDR, no slide deck.
Not ready to talk? See how we migrated Hapana off AWS →
Or call +61 2 8359 9507 · Hello@aviato.consulting