01
Problem
Psychological support scenarios require useful responses without losing risk controls, privacy boundaries or a traceable account of how multiple agents reached a result.
02
Role
System architect and primary engineer across agent coordination, runtime harness, retrieval, safety workflow, backend services and deployment.
03
Architecture
- 01
Coordinator-led event blackboard with claims and artifacts - 02
Understanding, Safety, Context and Response agents - 03
Redis short-term memory with durable database recovery - 04
Local Ollama and cloud model adapters behind one runtime
04
Key contributions
- Decoupled agent execution, persistence and post-processing in a reusable runtime harness.
- Implemented high-risk escalation, safety review and auditable report generation.
- Added dynamic SKILL.md loading and six MCP tools, including queued high-risk follow-up tasks.
- Deployed the static portal and application behind Caddy with HTTPS and restricted administrator routes.
05
Evaluation
- A versioned 60-query retrieval set and 23 automated test methods cover core runtime and safety paths.
- HitRate and MRR remain unpublished until the evaluation report is regenerated and independently reproducible.
06
Technology
07