AfterMath

Agentic post-disaster loss inventory. A policyholder uploads the messy photo pile from a fire/flood/storm; OpenCV 5 cross-examines every photo (sharpness, exposure, duplicate views, damage candidates, room grouping) and an agent loop turns those findings into decisions — rejecting unusable photos, demanding the missing shots, escalating contradictions to a human, and finally emitting a reconciled inventory with per-item confidence. The visual evidence provably changes what the agent does next: every round's perception digest and decision are recorded in a judge-readable trace.

Built for the OpenCV + AWS competition (Agentic Vision path). OpenCV 5.0.0 performs all substantive image analysis; the whole pipeline runs serverless on AWS (Lambda container image on Graviton/arm64 + S3 + DynamoDB) in region ap-southeast-2.

Quick start (local)

python -m venv .venv
.venv\Scripts\pip install -r requirements-dev.txt     # pinned: opencv-python==5.0.0.93, numpy==2.5.4
.venv\Scripts\python -m pytest tests                  # 15 tests: analysis ground truth + agent e2e
set "AFTERMATH_STORE=json" && .venv\Scripts\python -m app.api.local_server
# open http://127.0.0.1:8000

Generate the sample dataset and run the scripted evaluation (both reproducible, no API calls, ~30 s):

.venv\Scripts\python -m eval.make_dataset     # -> eval/dataset/*.jpg + manifest.json
.venv\Scripts\python -m eval.run_eval         # -> eval/results/results.{json,md}

Optional local smoke test of the real HTTP stack: .venv\Scripts\python smoke_local.py.

Deploy to AWS

Requires: AWS CLI v2 logged in (aws login). No Docker needed — the zip deploy stages aarch64 wheels directly:

# provider defaults to deterministic policy ("mock"); add an LLM for live demos:
powershell -File deploy\deploy-zip.ps1 -LlmProvider anthropic -LlmKey sk-ant-... -LlmModel claude-haiku-4-5

(On Linux/CloudShell: LLM_PROVIDER=openai LLM_KEY=... bash deploy/deploy.sh. A Docker-based variant deploy\deploy.ps1 builds the same thing as a container image via ECR.)

The script stages pinned aarch64 wheels (OpenCV 5 headless + numpy + requests), uploads the zip to S3, creates the S3 evidence bucket, the DynamoDB session table, a least-privilege IAM role, the Lambda function (python3.13, arm64/ Graviton), and an API Gateway HTTP API as the public endpoint — note that Function URLs with anonymous auth are blocked by this account type's managed policies, which is why the public entry is API Gateway. Current deployment: https://jw8tmot50b.execute-api.ap-southeast-2.amazonaws.com (verify with /healthz). Teardown: powershell -File deploy\teardown.ps1.

LLM provider options (AFTERMATH_LLM_PROVIDER): mock (deterministic policy — used by tests/eval), anthropic, openai, gemini (key via AFTERMATH_LLM_KEY, model via AFTERMATH_LLM_MODEL). The LLM receives the same structured perception digest as the policy and its output is validated against the live session; on any invalid or failed LLM response the loop falls back to the policy and records that in the trace — the demo cannot break because of a provider outage.

API

Method	Path	Purpose
GET	/healthz	liveness + OpenCV version
POST	/api/session	open a claim {incident_type, items:[{name,room,qty}], notes}
POST	/api/session/{id}/photo?name=&item=	upload photo (raw body) — runs full OpenCV 5 analysis
POST	/api/session/{id}/round	run one agent round (perception -> decision -> action)
POST	/api/session/{id}/attribute	tag a photo with {photo_id, item_id}
POST	/api/session/{id}/ack	mark a human-review escalation resolved
GET	/api/session/{id}	full session incl. rounds trace and report
GET	/api/session/{id}/thumb/{photo}	server-side thumbnail (cv2.resize)

Evaluation summary (policy decider, Windows dev box, see eval/results/)

8 scripted sessions with a simulated claimant: 7/7 counted sessions completed with all items confirmed (avg 2.4 rounds, avg 25 ms OpenCV analysis per photo); the 8th is a designed failure case (claimant never produces a usable photo) that correctly ends unresolved instead of inventing evidence. Duplicate detection, quality gates, quantity interrogation and contradiction escalation are all exercised and asserted in tests/ with known ground truth.