A Go-based AI multi-agent intelligent self-service help and support system for troubleshooting PostgreSQL and its derivative databases (like AlloyDB Omni) hosted on Kubernetes or VM infrastructure. The key features are:
- aiHelpDesk is an implementation of the shift-left support paradigm in AI-Assisted Database Management products (see next section).
- aiHelpDesk is designed for human operators as well as for the upstream agents, which are treated as the first class citizens.
- aiHelpDesk aims to prevent incidents based on active reasoning, not just help troubleshoot them after they occur.
- aiHelpDesk offers not only the reasoning layer specific to your database inventory, but it also features the actuation arm to optionally make the remediation changes to restore service or optimize its use.
- aiHelpDesk includes a comprehensive eight-module AI Governance framework with the tamper-proof audit as the cornerstone of safe, responsible and transparent remedial adjustments.
- aiHelpDesk features a built-in incident diagnostic bundle management for vendor support.
- aiHelpDesk features a built-in fault injection framework.
- aiHelpDesk is implemented using Google ADK (Agent Development Kit) for Go and the A2A (Agent-to-Agent) protocol for modularity and extensibility where self-contained expert agents can be added or swapped from a Marketplace in favor to those shipped with aiHelpDesk out of the box.
aiHelpDesk is designed to help customers and agents with the AI-assisted triage, root cause analysis and remediation of database related problems on K8s and VMs. For the upstream agents, including agentic full-featured SRE systems, aiHelpDesk can be thought of as the database reasoning layer that aggregates the myriad of signals emitted by a database (statistics, metrics, wait events, logs, traces, etc.) into a coherent story and offers a way to repair any of the problems found.
While SaaS applications clearly have their market and cloud vendor DBaaS systems in particular are among the fastest growing and most profitable sectors on GCP, AWS and Azure, there are legitimate reasons for many customers to stay away from the black-box, vendor lock-in, cloud provider specific management systems. See here for extensive treatment of this topic and, in particular, check out the 13 specific customer expectations of the cloud provider's DBaaS and how the actual cloud offerings mostly fall short of these expectations.
Enter the world of AI-Assisted Database Management products.
aiHelpDesk is the first product from the DDS Group that starts on the path of implementing the AI-Assisted vision of this new breed of the database management products.
Before diving into aiHelpDesk we recommend reviewing our design principles and the FAQ that may address some of the questions about the product.
aiHelpDesk can be deployed on K8s or VMs / bare metal. The binary packages are provided for Linux x86-64 and ARM (Graviton, Ampere), as well as for macOS (Intel and Apple Silicon).
There are two options to run aiHelpDesk on non-K8s environments, either in the Docker containers or straight on the host.
For the former, the first Docker Compose command brings up all the aiHelpDesk agents. The second command starts an interactive session of the aiHelpDesk Orchestrator to talk to a human operator:
docker compose -f deploy/docker-compose/docker-compose.yaml up -d
docker compose -f deploy/docker-compose/docker-compose.yaml --profile interactive run orchestrator
For the latter (i.e. for running aiHelpDesk with no Docker, straight on a host), there's a small helper startall.sh script that brings up all the agent Go binaries in the background including the optional Gateway, followed on with the interfactive aiHelpDesk Orchestrator session as well:
./startall.sh
Please be sure to set your desired LLM model (it defaults to Anthropic's Haiku), the API key and the database inventory.
See VM-based Deployment for detailed instructions on how to deploy aiHelpDesk either as binaries (simpler) or manually by cloning the repo.
tar xzf helpdesk-v0.1.0-deploy.tar.gz
kubectl create secret generic helpdesk-api-key --from-literal=api-key=<YOUR_API_KEY>
helm install helpdesk ./helpdesk-v0.1.0-deploy/helm/helpdesk \
--set model.vendor=anthropic \
--set model.name=claude-haiku-4-5-20251001
See K8s-based Deployment for detailed instructions on how to deploy aiHelpDesk on K8s.
See aiHelpDesk's ARCHITECTURE.md for system design, configuration, and extension guide.
aiHelpDesk is proud to feature a sophisticated AI Governanance system, which rests on eight separate subsystems, including full auditing. Compliance Reporting — the periodic governance and security posture assessment — is documented separately in COMPLIANCE.md and so are the aiHelpDesk Journeys, etc. Please see here for details.
HTTP-level authorization (who can call which endpoint, role definitions, role aliases) is documented in AUTHZ.md. Identity provider configuration (static users file, JWT/OIDC, service accounts) is in IDENTITY.md.
aiHelpDesk features a comprehensive testing strategy as documented here, including a built-in fault injection testing framework, see here.
In addition to the interactive Orchestrator, aiHelpDesk provides a Gateway REST API for programmatic access:
# Query the system
curl -X POST http://localhost:8080/api/v1/query \
-H "Content-Type: application/json" \
-d '{"agent": "database", "message": "What is the server uptime?"}'
# List agents
curl http://localhost:8080/api/v1/agents
# List managed databases
curl http://localhost:8080/api/v1/databases
# Direct tool calls
curl -X POST http://localhost:8080/api/v1/db/get_server_info \
-H "Content-Type: application/json" \
-d '{"connection_string": "host=db.example.com port=5432 dbname=mydb user=admin"}'The Gateway API documents the full REST API reference: all 17 endpoints with request/response shapes, query parameters, and curl examples. It is recommended for CI/CD pipelines, automation, and containerized environments. See deployment READMEs for details: for Docker, for running directly on a host or for running on K8s.
aiHelpDesk includes helper scripts for working around the ADK REPL bug in containers:
| Script | Description |
|---|---|
scripts/gateway-repl.sh |
Interactive REPL using the Gateway API (recommended for containers) |
scripts/k8s-local-repl.sh |
Run orchestrator locally with K8s agents port-forwarded |
See scripts/README.md for detailed usage.
aiHelpDesk can certainly be used by humans and that's what the interactive LLM-powered Orchestrator is there for. Additionally however, an upstream agent or a program can call aiHelpDesk agents directly as well. Here are some examples:
See a sample of a O11y watcher program or an SRE bot that calls aiHelpDesk (via a Gateway) to understand a state of a database and ask for AI-powered diagnostic and troubleshooting.
See a sample of a Security Responder bot that automatically sends alerts in real-time for security violations (e.g. for low confidence agent delegations, chain tampering, off-hours activity, high error rates, unauthorized destructive operations, etc.) and optionally creates a security incident (with the full incident bundle snapshot).
See a sample of a Compliance Reporter bot that queries the aiHelpDesk Gateway's governance API endpoints and produces a structured compliance snapshot. It is designed to run on-demand or on a schedule (e.g. daily cron / Kubernetes CronJob) and optionally post a summary to a Slack webhook. In contrast to SEC bot (reactive, threat-driven) and Auditor (streaming, rule-based alerts), GOV bot is designed to be periodic and analytical — the compliance officer's tool rather than the on-call engineer's troubleshooter.
See a Real-Time Auditor that can be used as an inspiration for an upstream long-running agent. In constrast to the SRE bot and the GOV bot that can be considered as one-shot automation agents, the SEC bot and the Auditor can be viewed as the core daemons. Indeed, both are long-running and both connect to the audit Unix socket and process events in real time. The distinction is purely in what they do: the Auditor fires webhook/email/syslog alerts, while SEC bot creates an incident bundle (via aiHelpDesk Gateway).
aiHelpDesk is designed to work with humans and upstream agents alike. Here's a sample intro dialog with a human operator (aka aiHelpDesk's "Hello World").
aiHelpDesk relies on Google ADK (Agent Development Kit) for Go, which was built around Gemini models. We've extended aiHelpDesk to work with both Anthropic and Gemini models.
Supported models:
| Vendor | Model Name | Notes |
|---|---|---|
| Anthropic | claude-haiku-4-5-20251001 |
Fast, cost-effective |
| Anthropic | claude-sonnet-4-20250514 |
Balanced performance |
| Anthropic | claude-opus-4-5-20251101 |
Most capable |
| Gemini | gemini-2.5-flash |
Fast, recommended for most use cases |
| Gemini | gemini-2.5-flash-lite |
Fastest, lower cost |
| Gemini | gemini-2.5-pro |
Most capable 2.5 model |
| Gemini | gemini-3-flash-preview |
Latest 3.0 series, fast |
| Gemini | gemini-3-pro-preview |
Latest 3.0 series, most capable |
Note: Gemini 1.x and 2.0 models are retired and will return errors.
Beyond these, support for other models can be easily added (ADK's LLM interface is simple and can be implemented for other providers, just as we did for Anthropic).
Please note that aiHelpDesk offers the flexibility for individual expert agents (e.g. a Database agent, a K8s agent, an Incident Management agent) to run with different LLMs if needed or if an agent's provider recommends or tests their agent with a particular LLM. The sample deployment scripts assumes the same LLM for all agents, but that can be easily adjusted with setting env variables before starting each agent.
Please contact aiHelpDesk if you or your customer would like to see a support for a particular LLM.