AssessAuto

Multi-agent vehicle-damage assessment. A LangGraph supervisor reasons over MCP-wrapped models and coordinates specialist agents through a full detect → describe → severity → price → fraud chain — with typed hand-offs and replayable trajectory observability.

This repository showcases the core agentic workflow: the supervisor, the MCP tool boundary, the vision-reasoning and specialist agents, and the chat app. It is a portfolio build replicating a multi-agent architecture from my industry PhD research.

Architecture

flowchart TB
    user([User query + vehicle image]) --> sup
    subgraph client["Supervisor process — Claude/Bedrock brain + MCP client"]
        sup["Supervisor<br/>ReAct: routing, orchestration, synthesis"]
        mem["Memory<br/>thread checkpointer + DynamoDB"]
        sup <--> mem
    end
    sup <-->|MCP tool calls| mcp
    subgraph server["AssessAuto MCP server — the model-serving boundary"]
        mcp["MCP server (5 tools)"]
        mcp -->|SageMaker| det["detect_damage · GDINO+SAM detector"]
        mcp -->|Bedrock| vis["describe_damage · Claude"]
        mcp -->|Bedrock| sev["score_severity · Claude"]
        mcp -->|Bedrock| price["estimate_price · Claude"]
        mcp -->|Bedrock| fraud["flag_fraud · Claude"]
    end
    sup --> ans([Unified answer + trajectory])
    sup -.trace.-> log["OTel spans / replayable trajectory"]

Loading

The supervisor is a pure MCP client — it never calls the models directly. Every model capability is an MCP tool. The damage detector runs as a separate Amazon SageMaker endpoint (invoked by detect_damage); the reasoning agents run on Claude via Amazon Bedrock. Each hand-off is a validated, typed structure (never free text), and every run emits a replayable trajectory.

How it works

1. detect_damage — finds damage regions (boxes + labels) via the SageMaker detector.
2. describe_damage — annotates the regions and asks Claude for a typed per-region assessment (type, location, severity, description).
3. score_severity — normalises the assessment into consistent 0–100 scores.
4. estimate_price — an AUD repair-cost range with a repair-vs-replace call.
5. flag_fraud — cross-checks detector regions against the description for inconsistencies.

The supervisor runs a ReAct loop and uses only the tools a given question needs — a narrow "what damage is visible?" stops after describe_damage; a full assessment runs the whole chain — then reconciles everything into one answer.

Sample output

A real run on a collision photo: the detector boxes the shattered windshield, and the agents produce a full assessment — overall severity, an AUD repair estimate with a repair-vs-replace call, and a consistency check — all with an expandable agent trajectory.

Detection — the detector boxes the damage.	Assessment — severity, AUD estimate, consistency check.

Sample vehicle photo from the public CarDD dataset.

Run it

Prerequisites: Python 3.12, AWS credentials with Amazon Bedrock model access, and a running SageMaker detector endpoint (set its name in .env).

python -m venv .venv && source .venv/bin/activate   # Windows: .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp .env.example .env                                 # set AWS_REGION, DETECTOR_ENDPOINT_NAME, BEDROCK_MODEL_ID

Then, in two terminals:

python mcp-server/server.py                          # 1) the MCP server (5 tools)
streamlit run app/streamlit_app.py                   # 2) the chat interface

Project layout

config.py        # one typed view over the environment (secrets from env only)
supervisor/      # the LangGraph brain: graph, MCP client, specialists, tracing
vision/          # annotate + schema + Bedrock reasoning (the Describer)
mcp-server/      # the MCP server exposing the five tools
memory/          # thread checkpointer + DynamoDB (dual memory)
app/             # Streamlit chat interface

Notes

• The SageMaker detector and Bedrock require your own AWS account and enabled model access.
• Secrets are read from the environment only — .env is git-ignored; never commit real credentials.

License

MIT — see LICENSE.