Imported from roshaninfordham/brainauthai (docs/AGENTS.md). Install upstream with npx skills add roshaninfordham/brainauthai --skill docs. Copyright stays with the author.

Agents

BrainAuth AI uses deterministic agents for the hackathon MVP. Each agent emits visible action traces: tool call, evidence found, confidence, and verification status. Hidden chain-of-thought is never displayed.

Triage Extractor Agent

• Input: EHR note
• Tools: parse_ehr_note, verify_grounding
• Output: age, NIHSS, last-known-well, symptoms, glucose, blood pressure, baseline mRS
• Failure behavior: missing facts become documentation gaps or human-review items

Imaging Evidence Agent

• Input: CT/CTA report
• Tools: analyze_cta_report, verify_imaging_sources
• Output: LVO, vessel territory, hemorrhage exclusion, ASPECTS
• Failure behavior: low-confidence imaging evidence requires human review

Policy Criteria Agent

• Input: synthetic payer policy
• Tools: read_payer_policy, prepare_criteria_matrix
• Output: normalized documentation criteria
• Failure behavior: ambiguous criteria are marked needs review

Medical Necessity Agent

• Input: clinical facts and policy criteria
• Tools: match_policy_criteria, verify_grounding
• Output: criteria documentation view and medical-necessity draft inputs
• Failure behavior: unsupported claims are blocked

Gap Detector Agent

• Input: facts, criteria, transfer note
• Tools: detect_documentation_gaps, prepare_gap_checklist
• Output: owner-based missing documentation checklist
• Failure behavior: missing values are not invented

Packet Builder Agent

• Input: facts, criteria, gaps, audit events
• Tools: build_packet, verify_packet_caveat
• Output: clinician-review packet draft, medical necessity draft, FHIR-style JSON
• Failure behavior: packet export remains draft-only and requires physician signoff

Audit Verification Agent

• Input: all agent outputs
• Tools: verify_source_coverage, verify_review_flags, verify_grounding
• Output: unsupported claims count, evidence coverage, human-review items
• Failure behavior: unsupported claims prevent clean export status

B.E. FAST Video Intake Agent

• Input: uploaded video or synthetic demo clip metadata
• Tools: sample_video_frames, check_frame_quality
• Output: representative frame set, stream readiness status, source clip metadata
• Failure behavior: insufficient video quality becomes a review item; no diagnosis is made

Azure Vision Frame Agent

• Input: sampled video frames
• Tools: analyze_be_fast_frames
• Output: visual findings for balance, eye/vision cue, face asymmetry, arm drift, and speech cue proxy
• Failure behavior: missing or low-confidence signals are marked needs review; Azure is never shown as connected unless credentials exist

B.E. FAST Classifier Agent

• Input: visual findings and B.E. FAST signal definitions
• Tools: score_be_fast_signals, apply_time_action_rule
• Output: B.E. FAST alert score and signal-level confidence bars
• Failure behavior: uncertain cases prompt human/caregiver review; sudden symptoms still trigger emergency-first language

EHR Fusion Agent

• Input: B.E. FAST findings and longitudinal synthetic EHR context
• Tools: fuse_ehr_risk_context, verify_context_sources
• Output: risk-context summary for monitoring escalation
• Failure behavior: missing EHR context is shown as missing context, not inferred

Escalation Agent

• Input: B.E. FAST alert score, EHR fusion output, care-circle contacts
• Tools: prepare_emergency_alert, notify_care_circle, request_clinician_callback
• Output: emergency prompt, loved-one notification, urgent stroke clinician callback request
• Failure behavior: emergency services remain primary; routine appointment language is not used for sudden stroke symptoms

Monitoring Audit Agent

• Input: video events, B.E. FAST signals, EHR context, escalation actions
• Tools: verify_monitoring_safety, log_monitoring_actions
• Output: monitoring audit trail with no diagnostic claim
• Failure behavior: any unsupported monitoring claim is blocked or marked for human review

Monitoring Agent Flow

flowchart LR
  Video[Upload or Demo Video] --> Intake[B.E. FAST Video Intake Agent]
  Intake --> Vision[Azure Vision Frame Agent]
  Vision --> Classifier[B.E. FAST Classifier Agent]
  Classifier --> Fusion[EHR Fusion Agent]
  Fusion --> Escalation[Escalation Agent]
  Escalation --> Audit[Monitoring Audit Agent]
  Audit --> Human[Emergency-first human response]