A Python and Streamlit research demonstrator for exploring saved synthetic-data experiments in vessel anomalies, drone observations, and cybersecurity classification.

Status: experimental. Not validated for operational decisions or production use. The dashboard reads committed CSVs and PNGs; it does not run inference, receive live feeds, or control equipment.

Streamlit research demo · Command-centre concept (Lovable) · Installation · Architecture · Limitations · Security

The hosted demo may sleep. This repository does not establish which revision is currently deployed.

Explore the coastal command-centre concept built with Lovable

This separate prototype explores citizen reporting, jurisdiction-based staff views, geospatial situational awareness and AI-assisted briefings with human review. The long-term ambition is a national, state, district and station hierarchy. National-scale capacity, operational readiness and official endorsement have not been established by this repository.

The public landing and reporting pages can be viewed without an account; staff views require authorised access. The linked site's reporting forms are not a sandbox verified by this repository. Use the link for interface review only: do not submit real incidents, personal details or restricted information during a demonstration.

The Lovable application is hosted and maintained separately. Its source, deployment and database are not included in this checkout, and it is not connected to the Python models or benchmark. Linking the two projects does not establish an implemented integration. See the architecture relationship and proposed integration.

LOF, Logistic Regression, SVM, live AIS, geospatial replay, autonomous surveillance, and AI-assisted incident analysis are not implemented on this baseline. The separate ICCC platform proposal in PR #10 is unmerged work and is not part of this demo.

Use Python 3.13 and uv 0.13.1.

git clone https://github.com/bikramtheitguy/ai-coastal-security-defense.git
cd ai-coastal-security-defense
uv sync --locked --no-dev
uv run --locked --no-dev streamlit run dashboard/app.py

Open the local URL printed by Streamlit. No credentials or external data services are required. See the pip-only instructions for Windows and Linux, and the explicit Community Cloud settings.

Do not activate the committed venv/ directory. It is a historical Linux/Python environment retained with the original repository; supported installations use a fresh .venv/.

Class 1 is the positive class. These figures describe the committed prediction files, not independently reproduced model runs.

| Experiment | Rows | TP | FP | FN | TN | Precision | Recall | F1 | Accuracy | Always-negative accuracy | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | Drone | 200 | 2 | 8 | 14 | 176 | 20.0% | 12.5% | 0.154 | 89.0% | 92.0% | | Cyber | 200 | 0 | 1 | 4 | 195 | 0.0% | 0.0% | 0.000 | 97.5% | 98.0% |

Both saved classifiers fall below an always-negative accuracy baseline. Accuracy alone obscures missed positives. The cyber sample has only four positive cases; these results cannot estimate operational reliability.

The artifact manifest records SHA-256 digests of the original CSVs and charts. Tests verify integrity and metric consistency. Original training split IDs, scores, labeling rules, generator code and seeds were not saved. Checksums do not recover that provenance.

A separate vessel-speed benchmark now generates documented synthetic observations, fits on training data only, selects a model and threshold using validation data, and evaluates a disjoint test period. It reports false alerts, missed anomalies, ranking metrics and baselines, and records IDs, scores, source hashes and runtime versions.

uv sync --locked --group research
uv run --locked --group research python -m coastal_security.vessel --output artifacts/vessel-demo
uv run --locked --group research pytest training_tests

Use a new output directory for each run. This benchmark does not replace the historical dashboard plot or establish real-world detection quality. Slow ordinary and anomalous scenarios intentionally overlap, illustrating the limits of speed-only detection.

dashboard/               Streamlit viewer, result helpers, saved artifacts and manifest
coastal_security/vessel/  New synthetic generator, training, evaluation and experiment CLI
data/                    Original datasets described as synthetic
models/                  Original serialized research models; never loaded by the demo
notebooks/               Legacy experiments and notebook outputs; known limitations remain
tests/                   Demo, data-contract, integrity and setup tests
training_tests/          Vessel benchmark isolation, metrics and reproducibility tests
docs/                    Installation, architecture, limitations and provenance
.github/workflows/       Demo checks and dependency auditing
pyproject.toml            Runtime, development and optional research dependency groups
uv.lock                  Universal dependency lock
requirements.txt         Hashed runtime export for pip and Streamlit Community Cloud
requirements-research.txt Hashed optional research export
venv/                    Historical environment; retained, unsupported and unused

uv sync --locked
uv run --locked ruff check .
uv run --locked ruff format --check .
uv run --locked pytest

CI checks the demo and new vessel benchmark on Windows and Linux with Python 3.13, verifies dependency exports, and audits runtime and research dependencies. Passing checks establish software behavior on synthetic examples, not operational model validity. See verification scope.

1. Extend vessel evaluation to approved representative data with predefined acceptance criteria.
2. Add grouped temporal evaluation for vessel trajectories and contextual geospatial features.
3. Introduce reproducible drone/cyber pipelines with documented data provenance and saved preprocessing.
4. Add geospatial replay and approved-source adapters.
5. Evaluate genuine computer vision and event-monitoring pipelines.
6. Introduce evidence-linked AI assistance with human review.

Each milestone needs separate validation. No automatic enforcement or operational decisions are provided.

Project code is distributed under MIT. Dataset provenance remains incomplete; the code license is not a substitute for per-asset rights verification. See data provenance and third-party notices.