A research project exploring LLM-assisted trauma triage decision support. The agent takes free-text EMS trauma reports and determines a trauma activation level using a hybrid approach: deterministic evaluation for vital signs and LLM-based reasoning for mechanism-of-injury and anatomical criteria.
- Extraction — An LLM extracts structured fields (vitals, injuries, mechanisms) from the free-text report
- Deterministic Evaluation — Vital sign criteria are evaluated against age-appropriate thresholds
- LLM Evaluation — Mechanism-of-injury and anatomical injury criteria are evaluated with clinical reasoning
- Triage Decision — Results are merged and the highest activation level is reported
Activation levels: Level 1 (critical), Level 2 (high-priority), Level 3 (moderate), or Standard Triage (no criteria met).
SPEC.md— Full product specification (architecture, data flow, criteria definitions)IMPLEMENTATION.md— Technical implementation plan (file structure, component hierarchy, build order)
SvelteKit 2 + Svelte 5, Tailwind CSS 4, shadcn-svelte, Anthropic Claude (Haiku for extraction, Sonnet for evaluation), Zod for schema validation.
pnpm install
pnpm devRequires an ANTHROPIC_API_KEY environment variable (see .env.example).
For live speech-to-text input, also set DEEPGRAM_API_KEY.