RAG
Responses grounded in vetted gold standard content — no open-web fallback
B.R.I.D.G.E.
Rule-based governance layer — validate, repair, deliver
Zero-training
No custom model training required — deployable as system augmentation
Context & Problem
The emotional stakes of survivorship AI are uniquely high
Lung cancer survivors face a distinctive burden — not only the clinical realities of post-treatment life, but persistent stigma rooted in smoking history, heightened anxiety, fear of recurrence, and limited access to specialists in many communities.
Standard conversational AI is poorly suited to this population. Tone misalignment, excessive questioning, ambiguous advice, and stigmatizing phrasing — however subtle — can increase distress, undermine trust, and create reputational and liability risk for the organizations that deploy them.
Prompt engineering alone cannot reliably prevent these failures. Prompt iteration outpaces safety validation, and most RAG implementations treat model output as trusted without independent verification. A structural solution was needed.
Product Screens — Prototype
Governed responses in a live context
Each response is grounded in curated survivorship content, validated by B.R.I.D.G.E., and delivered with stigma-sensitive, empathy-first framing — including multilingual support.
RAG-grounded tobacco response with B.R.I.D.G.E. comms check badge and source attribution
Spanish-language fatigue response — safety-validated before translation, with source and comms check
Empathy-first pain management response with avatar interface and module source citation
The B.R.I.D.G.E. Framework
Separating generation from governance
B.R.I.D.G.E. — Behavioral Rules for Informed Dialogue in Guided Engines — is an evaluation-driven governance layer that sits between generation and delivery. It executes independently of the base model, ensuring policy enforcement is never dependent on model compliance alone.
Core principle: Every response is generated, then validated against a formal rule set, then repaired if violations are detected — before anything reaches the user. The model generates; B.R.I.D.G.E. governs.
R1
Empathy Reflection
Ensures emotional acknowledgment before informational content. Hybrid architecture: deterministic insertion paired with LLM refinement for naturalness.
R2
Anti-Stigma Enforcement
Deterministic phrase replacement using longest-match-first logic. Flags and rewrites stigmatizing language before delivery, including smoking-related framing.
R3
Open-Question Limits
Structural edit enforcing caps on interrogative density. Prevents cognitive overload through heuristic and regex-based detection.
R4
Medical Framing Boundaries
Constrains responses to factual, non-prescriptive framing. No independent diagnosis, no directive clinical instructions.
R5
Agency Affirmation
Reinforces the user's autonomy and decision-making capacity. Guards against paternalistic or directive tone patterns.
R6–R8
Tone Calibration, Fallback Safety & Multilingual Validation
Supportive tone calibration; safe fallback responses when retrieval fails; safety validation applied before translation to preserve intent across languages.
Architecture
Retrieval-grounded generation with an independent governance layer
The Companion pulls exclusively from curated survivorship materials via a FAISS-based vector retrieval system. Generated responses pass through the B.R.I.D.G.E. rule engine — validate, repair, deliver — before reaching the user. Bounded repair loops prevent latency escalation.
Orchestration & Models
- Python + LangChain orchestration
- OpenAI and Anthropic model compatibility
- FAISS-based vector retrieval
- Bounded repair loops (latency-controlled)
Governance Layer
- B.R.I.D.G.E. rule engine (8 operational rules)
- Deterministic phrase replacement
- Structural edits (question limits, tone)
- Structured JSON violation logging
Knowledge & Content
- Curated survivorship materials only
- No open-web fallback
- Safe fallback-mode when retrieval fails
- Reading-level transformation support
Observability & Deployment
- Langfuse prompt version tracking
- Pre/post repair comparison logging
- No custom model training required
- Web-based, hotline-augmentation, or care-nav integration
Evaluation & Risk Controls
Evaluation embedded in the runtime, not bolted on after
Every rule in B.R.I.D.G.E. has a corresponding test suite. Guardrails can be toggled, traced, and regression-tested independently, so prompt iteration does not introduce unmeasured conversational risk.
| Control |
Purpose |
| Strict RAG gating |
All responses grounded in vetted survivorship content. No open-web speculation. |
| Deterministic rule enforcement |
Anti-stigma and framing rules are hard-coded — not subject to model judgment or drift. |
| Bounded repair loops |
Failed validations trigger repair, with loop limits to prevent latency escalation. |
| CSV-driven rule test suites |
Each rule has deterministic and soft-violation test cases scored with pass/fail metrics. |
| Advisory signal classification |
Qualitative rules emit advisory signals logged for SME review rather than hard blocking. |
| Pre/post repair logging |
Every repair is logged with original and repaired output for audit and regression analysis. |
| Prompt version tracking |
Langfuse integration traces prompt changes across deployments to isolate regression sources. |
| Multilingual safety validation |
Safety checks run on source language output before translation preserves intent across languages. |
Deployment & Vision
A tireless digital worker — built for the organizations that serve survivors
The Companion is designed as an augmentation tool, not a replacement for clinical care. It requires no custom model training and can be deployed as a web-based companion, a hotline support augmentation, or a digital extension of an existing care navigation program.
Target deployments include nonprofits, rural survivorship programs, and academic research centers where specialist access is limited and between-visit support is most needed.
Design principle: Always available, consistently supportive, and built to uphold the dignity and lived experience of lung cancer survivors — including communities where stigma has historically created barriers to care.
About
George L. Reynolds
Health Technology / Digital Health Product Leader with 20+ years of experience building and commercializing clinical-grade digital health products. The Lung Cancer Survivor Companion and B.R.I.D.G.E. framework were designed and developed by Health Decision Technologies in collaboration with the University of Colorado Anschutz Cancer Center, drawing on prior production AI work at FlyteHealth and NIH-funded mobile health research.
georgelreynolds.com
·
LinkedIn