Axxess Aegis

Curing clinician burnout by turning chaotic medical audio into audit-ready Axxess EMR records.

Comment

Inspiration

As a certified EMT-Basic working shifts at LifeLine EMS, I run rapid triage handoffs that reveal a critical flaw in healthcare tech: capturing accurate clinical data during chaotic moments is incredibly difficult. Missing vital details leads to severe compliance risks and compromised patient care. I built Axxess Aegis to solve this exact bottleneck, serving as an autonomous clinical shield designed specifically for the Axxess home health ecosystem.

What it does

Axxess Aegis is an AI diagnostic assistant that transforms raw clinical audio into structured data and actionable patient care.

  • Autonomous Extraction: It extracts vitals and symptoms while auto-tagging conditions with precise ICD-10 codes. It structures everything into a strict FHIR R4 JSON payload for immediate EMR syncing.
  • Agentic Gap Analysis: Aegis acts as a proactive clinical copilot. It evaluates transcripts against standard medical protocols and instantly flags missing diagnostic information.
  • Patient Accessibility: The platform translates complex medical jargon into highly understandable recovery plans. It features instant multilingual translation, a native Text-to-Speech module, and direct SMS delivery to the patient's mobile device.
  • Compliance First: A mandatory clinician verification gateway ensures audit-ready record generation and eliminates AI hallucination risks.

How I built it

I engineered a highly responsive full-stack application using Next.js and Tailwind CSS. I integrated the browser native Web Speech API for real-time audio capture. For my backend inference, I leveraged the Featherless API platform with the Mistral-Nemo-Instruct model. To handle my direct-to-patient communication pipeline, I integrated the Twilio SDK via secure serverless API routes.

Challenges I ran into

My primary hurdle was output token latency. Asking a massive language model to parse medical data, map ICD codes, generate follow-up questions, and format a deeply nested FHIR JSON payload caused severe processing delays of over 60 seconds. In a clinical setting, waiting a full minute for a screen to load is unacceptable.

I solved this with a three-pronged architecture rewrite:

  1. Prompt Flattening: I stopped forcing the LLM to write boilerplate schema and instead generated flat JSON objects, handling the complex FHIR mapping via JavaScript on the backend.
  2. Parallelization: I split the data extraction and patient translation into simultaneous API calls, maximizing my Featherless concurrency limits.
  3. UI Masking: I built a dynamic, multi-step loading state to psychologically mask the remaining wait time.

By shifting from sequential to concurrent processing, I fundamentally altered my latency equation from an additive bottleneck to a parallelized maximum: .

Accomplishments that I'm proud of

I successfully bridged the gap between a passive transcription app and a proactive AI agent. My autonomous gap analysis feature catches clinical mistakes in real time. I slashed a massive 68-second API bottleneck down to under 15 seconds. Finally, implementing Twilio allowed me to create a true end-to-end journey from clinical audio to a translated text message buzzing in the patient's pocket.

What I learned

I deepened my understanding of healthcare interoperability and the strict requirements of the FHIR standard. I mastered managing LLM concurrency limits and serverless AI infrastructure to achieve minimal latency. I also learned how to seamlessly integrate third-party communication SDKs like Twilio into a Next.js App Router architecture.

What's next for Aegis

My next step is upgrading my API mockup into a true SMART on FHIR integration. This will allow Aegis to securely authenticate and launch directly from within the Axxess EMR interface. I also plan to expand my insights engine to cross-reference longitudinal health records.

Built With

Share this project:

Updates