AskAivara: Clinical Reasoning Agent

• 

  Marketplace listing

Inspiration

Clinicians don't lack data. They lack synthesis. A patient can have two years of documented abuse, four consecutive years of depression screening with no referral, and a rising BMI trend in the obese range and still look unremarkable on a quick scan. AskAivara was built to find that patient and surface the full picture before the consultation starts.

What it does

AskAivara is an external A2A agent on the Prompt Opinion Marketplace. It reasons across a patient's full FHIR health record by chaining clinical data tools in sequence and returns a coherent clinical response covering patient summaries, risk assessments, and visit history. It identifies when specialist referrals are warranted and routes to them explicitly.

How we built it

Built with Google ADK using Gemini via Google AI Studio, deployed to a web server, and registered on Prompt Opinion as an external A2A agent. FHIR credentials are injected at runtime via the SHARP context extension before every model call. The agent is read-only and stateless across sessions.

Challenges we ran into

Getting the FHIR bearer token to flow correctly through the tool layer took longer than expected. Rather than continuing to debug the MCPToolset approach, we made a deliberate decision to replace it with native ADK tool functions that read credentials directly from session state. It was more work upfront but produced a cleaner and more reliable result.

Accomplishments that we're proud of

In a single consultation, AskAivara surfaces a coherent clinical picture across conditions, screening history, growth metrics, and clinical notes without the clinician having to open a single additional screen. The specialist routing logic correctly identifies when to escalate and when not to, which turned out to be harder to get right than the retrieval itself.

What we learned

Getting the credential boundary right early made everything downstream easier to trust. Runtime credential injection scoped to the session keeps the agent stateless and safe, and means any patient selected on the platform works correctly without any configuration change.

What's next for AskAivara: Clinical Reasoning Agent

More testing against real patient workspaces to make the routing logic more reliable across diverse data. Better handling of edge cases where the clinical picture is ambiguous. And exploring how AskAivara fits into larger multi-agent workflows as a specialist rather than always being the generalist entry point.

Built With

• a2a
• gemini
• googleadk
• python