scid-agent-showcase

A controllable LLM-agent workflow for high-risk structured clinical interviewing, packaged as a public engineering showcase with synthetic configs and replay examples.

Public showcase only. This repository is not a clinical diagnostic tool or medical device.

This repository highlights:

• Flow-controlled orchestration for structured interview phases and module progression
• Risk-aware interruption, event logging, and replayable session review
• A paper-backed system subset that stays inspectable without private data or unpublished assets

Demo: Watch workflow demo (60s) | Read architecture notes | Run the showcase

Overview

scid-agent-showcase is the public-facing subset of a larger research prototype for structured clinical interviewing. The code here is intentionally narrowed to the parts that are most useful for engineering review: interview flow control, schema-backed evidence collection, event logging, transcript replay, and compact report reconstruction.

Everything included in this repository is synthetic or safe to share. Private prompts, unpublished evaluation assets, real data, and internal experiment materials are intentionally excluded.

Interface

Core module dialogue	Structured screening

• The dialogue surface makes phase progression readable instead of hiding workflow decisions behind a generic chat shell.
• The screening view shows how structured yes/no turns map cleanly to controlled state transitions.
• These screenshots come from the runnable demo layer used in the portfolio and paper appendix workflow.

System Design

The public subset is centered around four inspectable areas:

• server/orchestrator/flow_controller.py: deterministic phase, module, and transition logic
• packages/schemas/: structured extraction contracts and JSON schema export
• server/orchestrator/event_bus.py and server/utils/logger.py: event history and structured logging hooks
• server/services/transcript_importer.py and server/services/report_service.py: replay import and compact report reconstruction

Evidence Map

Claim	Public evidence in this repo
Deterministic Flow Controller	server/orchestrator/flow_controller.py, configs/workflow.json, tests/test_showcase_flow.py
Schema-first I/O	packages/schemas/, configs/schemas/, tests/test_replay.py
Event logging and reviewable session state	server/orchestrator/event_bus.py, server/orchestrator/session_state.py, server/utils/logger.py
Transcript replay	server/services/transcript_importer.py, examples/transcripts/synthetic_case.md
Report reconstruction	server/services/report_service.py, examples/run_showcase.py

This public repository does not include a benchmark suite, unpublished evaluation assets, full prompt library, or real interview data from the larger research workspace.

Evaluation / Replay

• Risk-first checks can interrupt or skip unsafe branches before continuing the workflow.
• Session actions, event history, and structured logs make behavior reviewable across workflow versions.
• Replay and report reconstruction turn long dialogue traces into inspectable engineering artifacts instead of one-off demos.

What I Built

• Workflow and state-machine design for phase transitions, module activation, and failure handling
• Schema design for evidence collection, extraction contracts, and exportable JSON schemas
• Replay, logging, and report-packaging paths for inspectable end-to-end review
• Public showcase packaging, documentation, and runnable examples for hiring and collaboration review

Public Scope

Included here:

• Safe-to-share orchestration, schema, logging, and replay code
• Synthetic workflow definitions and toy transcript assets
• Lightweight tests and example runners for fast inspection

Intentionally excluded:

• Real or sensitive data
• Private prompts, keys, local environment files, and unpublished paper assets
• Large internal logs, unreviewed outputs, and deployment leftovers

Quickstart

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python -m packages.schemas.export --check
python -m pytest
python -m examples.run_showcase

Useful files:

• docs/architecture.md
• docs/public_release_manifest.md
• config.example.yaml
• CONTRIBUTING.md

Repository Guide

packages/schemas/        Schema models, registry, and schema export
server/orchestrator/     Flow controller, events, session state
server/services/         Workflow loading, question repo, replay/report helpers
server/utils/            Logging utilities
configs/                 Synthetic workflow, toy questions, JSON schemas
examples/                Replay transcript and demo runner
docs/                    Architecture and public release notes
tests/                   Minimal behavior checks for the showcase subset