How Ideas2IT Built the Agentic AI System That Turns Oncology Guideline PDFs into Deterministic Treatment Pathways for One of South Florida's Largest Public Health Systems

Guideline Structuring Pipeline: Converting PDF Pathways into Deterministic Decision Logic

The first engineering decision set the constraint for everything that followed: a general-purpose language model cannot safely traverse oncology guidelines that live as branching logic inside PDFs. The source had to be converted before inference was possible.

Ideas2IT built

1. a batch pipeline that takes MOFFIT guideline PDFs and converts them into structured decision-tree representations the inference layer can traverse with constraints.
2. All conditions, prior treatment checks, and biomarker references cited in each pathway were extracted upfront during this batch phase.
3. Pulling that extraction out of the inference loop reduced runtime complexity and made real-time recommendations faster and less error-prone.

The output of this pipeline is a structured pathway the agent can walk, step by step, with constraints enforced at each node.

This phase produced

• Batch PDF-to-decision-tree conversion pipeline
• Structured pathway representations of MOFFIT oncology guidelines
• Upfront extraction of conditions, prior checks, and biomarker references
• Constraint-encoded decision logic ready for deterministic retrieval

Deterministic Retrieval and Agentic Inference: An 8-Step Sequence That Mirrors Clinical Reasoning

With structured pathway logic in place, the design question was how the inference agent would retrieve context. Vector search was explicitly ruled out. Ideas2IT used entity matching instead, so context retrieval is predictable, reproducible, and auditable across every patient query.

On top of that retrieval layer, the inference agent runs an 8-step sequence that mirrors how an oncologist walks through eligibility and exclusions for each treatment line: performance status, prior drug exposure, resistance markers, drug family naming variants, comorbidities, and line-specific constraints are each addressed in sequence.

This mirrors clinical reasoning so the agent does not skip constraint checks the way a flat prompt would.

This phase produced:

• Entity-matching retrieval layer
• 8-step agentic inference sequence
• Performance status and prior exposure constraint handling
• Drug resistance and drug family naming logic
• Multi-line treatment pathway traversal
• Deterministic, auditable context retrieval

Reflective Validation Loop and Explainability Layer: Generation, Gap Detection, and Iterative Correction

A single-pass generative system is insufficient for clinical decision support. Ideas2IT implemented a reflective pattern using two connected agents: one generates the regimen shortlist with reasoning, and the second checks it for constraint misses, unsafe logical leaps, and gaps across the treatment lines covered.

Where gaps are found, the generation agent is forced to correct and resubmit. The cycle repeats until the output satisfies the validation rules.

The explainability layer runs through this same loop, producing reasoning for why each regimen was included and why others were rejected, traceable to the pathway constraints, not to probabilistic inference.

This phase produced:

• Generation agent for regimen shortlisting and reasoning
• Validation agent for gap detection and constraint verification
• Reflective correction loop
• Explainability output for regimen selection and rejection
• Guardrails against hallucination across multi-line treatment scenarios
• Support for inconsistent biomarker input formats