The AI System Helping Veterans Get Help When It Matters Most

A Houston-based nonprofit connects veterans and their families to a vetted network of partner organizations spanning employment, housing, healthcare, and other critical services. As their caseload grew, so did the complexity of determining which veterans needed immediate staff attention and which could be addressed through routine follow-up. The organization recognized that a data-driven, unbiased approach to triage could meaningfully improve outcomes for the veterans they serve. Rather than committing to a full production build, they chose to validate the concept through a focused proof of concept — using Generative AI to assess veteran history and surface the cases most urgently requiring human intervention, before making any larger infrastructure investment.

Protagona was engaged to design and deploy a working AI-powered priority scoring proof of concept on AWS. The system needed to ingest veteran connection data, apply a configurable severity algorithm, and output ranked priority tiers to guide staff intervention decisions. The POC had to demonstrate three things before any investment in a full production system: scoring accuracy across diverse veteran records, full transparency into AI reasoning, and an architecture sustainable enough for the client team to operate and evolve independently after handoff.

The core technical challenge was building a scoring system that could reason across a veteran's full history of service connections (not just a single intake field) and produce a defensible, auditable priority ranking. Real-world records frequently arrive with missing or incomplete fields, so the model had to handle data gaps gracefully without failing silently or dropping veterans from the queue entirely. Equally critical was making the AI's reasoning visible: staff needed to trust the output, which meant every scoring decision had to be explainable and cross-verifiable against source data in Salesforce.

‍

On the integration side, connecting an AWS Bedrock inference layer to an existing Typesense search infrastructure (already in use for labor market data queries) required dedicated research and close collaboration with both the organization's technical staff and frontline intake workers. Aligning the natural language processing approach with Typesense's query model introduced its own complexity, as did balancing token efficiency, inference cost, and scoring accuracy across a diverse veteran population when designing and iterating on prompts.

Protagona designed an event-driven, serverless architecture on AWS that routes veteran connection records through a Generative AI scoring pipeline. AWS Bedrock provides the large language model inference layer at the center of the system, receiving structured JSON payloads representing each veteran's full connection history. Lambda functions handle compute orchestration, API Gateway exposes a clean interface for triggering scoring requests from upstream systems, and DynamoDB stores results with hashing logic to detect new or changed records. S3 supports batch processing workflows for bulk record evaluation.

‍

Prompt engineering was iterative and collaborative. Working directly with intake staff, the team validated scoring criteria, refined priority weights, and incorporated secondary tie-breaker logic reflecting the organization's real-world triage judgment. Categorical data optimization reduced token consumption without sacrificing scoring fidelity. A specific edge case fix ensures that even records with missing data receive a lowest-priority designation rather than being silently dropped.

‍

Observability was built in from the start. CloudWatch captures the AI's reasoning narrative for every scoring decision, enabling staff to cross-verify outputs against Salesforce. Batch processing scripts and a structured testing framework supported user acceptance testing at scale, and knowledge transfer documentation was developed in parallel with the final sprint so the client team could operate and evolve the system independently after handoff.