Ask in plain language. Get the right hydrological simulation.
A unified, agentic, natural-language-driven platform for modelling water across river basins, cities, and countryside. Hydrological modelling today is fragmented and expert-gated — separate tools for catchments and for cities, each with its own data wrangling, setup, and interpretation. This platform puts one conversational interface in front of all of it: describe the question, and an agentic runtime assembles the data, selects and runs the right model, and explains the result — uniting data-driven and physically based approaches in a single workflow.
The platform is one agentic runtime conducting a set of specialised engines, designed to interoperate, over a shared open-data foundation. You speak to the runtime; it speaks to the models — and keeps every step as an inspectable, reproducible artifact.
The conductor. You describe a modelling goal in plain language; aiswmm coordinates the workflow while execution stays deterministic, inspectable, and artifact-based — with QA verification, provenance tracking, and modelling memory, so results are reproducible and auditable. It is exposed as MCP servers and Skills, so it also runs under Codex, Claude, Hermes, or OpenClaw. → agentic-swmm-workflow
A data-driven engine that learns rainfall–runoff behaviour from records and predicts river discharge and floods at catchment to regional scale. Best where gauged or large-sample data exist and you need fast streamflow and flood forecasts. → this repository (details below)
The SWMM framework the aiswmm runtime drives (same repository as above): a verification-first, auditable workflow for EPA SWMM — GIS preprocessing, model generation, deterministic SWMM runs, QA checks, calibration support, and provenance, natural-language driven but artifact-based, with the modeller in control. For urban & rural drainage, runoff, and Low-Impact-Development (LID) design. → agentic-swmm-workflow
A headless, natural-language automation layer for DHI MIKE+: an MCP server + Skills that let an agent inspect, edit, run, read, and plot a MIKE+ model end to end — no GUI. For detailed urban-water and flood modelling and batch/scenario automation; reading results and plotting need no MIKE+ license. → Agentic-MIKE-Plus
Draw an area (today, anywhere in Canada) and it pulls the open data for that spot — rainfall, terrain, land cover, soil, and the city's storm network — and assembles a complete, ready-to-run model. It uses real municipal networks where published (Victoria, Ottawa) and synthesises one elsewhere. This is the upstream data-preprocessing stage that feeds the engines above. → SWMMCanada
This repo is the platform's deep-learning rainfall–runoff engine: it trains LSTM networks to predict river discharge and floods at catchment scale. It is built to be read and learned from — you start on a zero-download synthetic basin and move to real catchment data, reusing the same model and training code throughout.
python -m venv .venv && source .venv/bin/activate
pip install -e .
python scripts/01_generate_data.py # synthetic rainfall–runoff
python scripts/03_train_lstm.py # train + evaluate (NSE ≈ 0.87 on held-out years)Why LSTMs? A river's flow today depends on weeks to months of past weather — snowpack, soil moisture, groundwater. An LSTM's memory is built to capture exactly that kind of long-range dependence, which is what makes it strong at streamflow and flood forecasting.
Learn the ideas: LSTM intuition · from LLMs to LSTMs · hydrology metrics.
MIT