Official Website: https://aria-project-homepage.web.app/

Inspiration

Coronary CT acquisition planning often depends on expert knowledge that is difficult to standardize, explain, and transfer. Radiologic technologists must consider heart rate, heart-rate variability, motion, phase strategy, temporal resolution, and scanner constraints under time pressure.

ARIA was created to make that decision process more structured, reproducible, and explainable without replacing the medical professional. For OpenAI Build Week, I extended the existing ARIA v1.9.1 system with Codex and GPT-5.6 to explore a clear separation between deterministic decision-making and conversational explanation.

What it does

ARIA v1.9.1 Build Week Edition is a non-diagnostic coronary CT acquisition support AI for radiologic technologists.

The Main workflow generates the formal acquisition strategy using deterministic logic. It evaluates the synthetic case, compares candidate strategies, and presents:

  • the confirmed rotation and beat strategy
  • acquired temporal resolution
  • reference heart-rate range
  • phase strategy
  • Candidate Comparison
  • Selection Rationale
  • Decision Trace

The formal result remains authoritative.

When the user needs an explanation, the current formal result, graph context, and Decision Trace can be handed off to Buddy Mode. GPT-5.6 then explains the graph and the selected strategy conversationally while using the formal result as read-only context. GPT-5.6 does not recalculate or replace the deterministic recommendation.

The application also includes:

  • English and Japanese interface support
  • separate Main and Buddy conversation histories
  • real-time streaming responses
  • an offline synthetic review workflow
  • an optional live GPT-5.6 workflow
  • a portable Windows review build

The demonstration uses a fully synthetic case and contains no patient data.

How we built it

ARIA Build Week Edition was developed as an independent extension of the existing ARIA v1.9.1 desktop application.

Codex was used throughout the project to:

  • extend the desktop UI and conversation workflows
  • implement Main-to-Buddy handoff
  • add structured graph context and Decision Trace transfer
  • integrate GPT-5.6 through the OpenAI Responses API
  • build bilingual interface and response controls
  • create a secure Local Relay workflow
  • implement temporary API-key handling
  • add automated regression, routing, packaging, and security tests
  • prepare the portable Windows review candidate
  • support video production and release verification

The application uses a strict authority boundary:

  1. deterministic logic generates the formal result
  2. GPT-5.6 receives that result as read-only context
  3. Buddy Mode explains the result
  4. the formal recommendation remains unchanged

For live use, the Windows application communicates with a Local Relay bound only to localhost. The OpenAI API key is entered temporarily, passed only to the owned child process, and is not stored in the ZIP, configuration files, command line, application logs, or user interface.

Challenges

The most important challenge was preventing the language model from becoming an accidental decision engine.

It was not enough to produce a helpful answer. The system had to preserve a visible and testable boundary between:

  • formal acquisition strategy
  • educational explanation
  • conversational interaction

We also had to prevent duplicate requests during automatic Main-to-Buddy handoff, keep the formal result unchanged after conversation, and safely package a multi-process desktop workflow for Windows.

Other challenges included:

  • maintaining consistent behavior across English and Japanese
  • securely handling optional API keys
  • keeping Local Relay traffic loopback-only
  • creating a portable one-folder review build
  • preserving deterministic output across UI changes
  • producing a complete demo in under three minutes

Accomplishments

During Build Week, ARIA gained:

  • GPT-5.6-powered Buddy explanations
  • explicit Main-to-Buddy handoff
  • structured read-only graph context
  • real-time streamed responses
  • bilingual workflows
  • a secure Local Relay
  • temporary BYOK support without key persistence
  • offline synthetic review capability
  • portable Windows packaging
  • extensive automated regression and security testing

For the demonstrated fully synthetic case, ARIA confirms:

  • heart rate: 62 bpm
  • heart-rate range: 61–63 bpm
  • PQ: 180 ms
  • motion assessment: Mild
  • confirmed strategy: 0.35-second rotation / 2 beat
  • acquired temporal resolution: 92.9 ms
  • phase strategy: Diastolic only

Buddy Mode explains this result without changing it.

What we learned

The strongest use of an LLM in this workflow was not to replace deterministic logic, but to make that logic understandable.

A reliable professional tool can benefit from two clearly separated layers:

  • a deterministic layer that owns the formal decision
  • a conversational layer that improves explanation, education, and usability

Codex was especially valuable for maintaining consistency across a large, evolving codebase, generating tests, examining routing behavior, and helping preserve safety boundaries during rapid development.

What's next

Future work includes:

  • a production-grade hosted relay option
  • expanded educational explanations
  • additional coronary CT workflow scenarios
  • improved multilingual documentation
  • broader external Windows testing
  • a Japanese-subtitled official tutorial
  • integration of the Build Week improvements back into the main ARIA project after review

ARIA will remain non-diagnostic, and final professional judgment will continue to belong to the radiologic technologist.

Built With

  • api
  • bilingual
  • codex
  • fastapi
  • gpt-5.6
  • local
  • openai
  • pyinstaller
  • python
  • relay
  • responses
  • streamlit
  • uvicorn
  • windows
Share this project:

Updates