Ideas2IT designed the solution as an agentic decision workflow with deterministic context retrieval and explicit step-by-step reasoning.
1) Built two separate pipelines for safety and speed
- Batch pipeline: Converts guideline PDFs into structured decision logic (decision-tree-like representation).
- Real-time pipeline: Takes pathology and biomarker inputs and returns regimen recommendations with explanations.
This separation ensured guideline processing doesn’t slow down inference and reduces runtime complexity.
2) Chose a model + platform built for production reasoning
- Implemented in Azure AI Foundry using GPT 4.1 to balance reasoning quality and latency.
3) Replaced vector search with deterministic retrieval
Instead of vector search, the system uses entity matching so guideline context retrieval is predictable and reproducible.
This was a deliberate design choice to reduce ambiguity and make behavior auditable.
4) Pre-extracted pathway constraints to reduce inference load
All conditions, prior checks, and biomarkers referenced in the guideline pathways were extracted upfront so real-time inference is faster and less error-prone.
5) Designed an 8-step inference agent to mimic clinician workflow
The inference agent follows an 8-step sequence that mirrors how an oncologist walks through eligibility and exclusions so it does not miss constraints across lines of treatment.
6) Added a reflective “generation + validation” loop
A connected pair of agents runs in a reflective pattern:
- one generates the regimen shortlist and reasoning
- the other checks gaps, constraint misses, and unsafe leaps
- then forces iterative correction until the output meets rules
