TRID3NT

Storms, tsunamis, earthquakes, (and more) simulated from a sentence.

Comment

Inspiration

The inspiration of TRID3NT came from the friction of doing real environmental and hazard modeling. The tools are powerful but locked behind specialist GIS software, hand-built model decks, gigabytes of data wrangling, and expensive compute -- a single flood or groundwater study can take a specialist days to weeks just to set up. We wanted to collapse that into a conversation: ask a question in plain language and get a rigorous, simulated answer on a live map.

What it does

TRID3NT removes the friction of GIS and environmental modeling systems with AI agents. It leverages cloud computation to deliver answers quickly without the need of expensive hardware or slow simulations. Its goal is to help experts streamline their workflow so they can focus more on solving problems and less on setting up simulations and data visualization.

TRID3NT uses the same models and data in use by USACE, FEMA, NOAA, and NASA, with many of them authored by USGS or other government agencies. The app can run complex hydrology (SWAN, SWMM), groundwater (MODFLOW), seismic (OpenQuake), and storm (SFINCS, GeoClaw) simulations. It can fetch satellite data and other datasets needed to model anything, and exposes 175+ composable geospatial tools including the full QGIS Processing toolbox that the agent calls on demand.

How we built it

  • A React + MapLibre web app on Vercel: a chat next to a live, interactive map.
  • The brain is an LLM agent on AWS Bedrock (Claude Sonnet/Haiku, Amazon Nova) driving 175+ tools over a WebSocket: data fetchers, QGIS/GDAL geoprocessing, ML inference, and physics-solver dispatch.
  • The engines are real, peer-reviewed numerical models (SFINCS, MODFLOW, PySWMM, OpenQuake, GeoClaw, SWAN) plus ML models (e.g. Meta's canopy-height network), containerized and run on AWS Batch (Spot) that scales from zero.
  • Results become Cloud-Optimized GeoTIFFs in S3, served as map tiles by TiTiler and drawn in MapLibre with legends and time animation.
  • The whole stack is scale-to-zero: the agent server auto-stops when idle and wakes on demand, Batch scales to zero between jobs, DynamoDB holds state, and serverless Lambdas serve saved results even when the agent is asleep.

Challenges we ran into

  • Cost vs. capability: real solvers are expensive, so everything had to scale to zero -- the only always-on piece is a tiny tile server.
  • Reliability over WebSocket through CloudFront (heartbeats, reconnect, per-case layer durability) after dropped connections.
  • Driving real solvers and the QGIS toolbox with an LLM: every engine has its own deck format; we wrapped them in tool contracts the agent can compose, with honest failure modes -- it degrades and tells you, it never fakes a result.
  • Letting the user supply input the agent cannot (drawing an area of interest, a barrier, or a point) -- so we built an on-map vector-drawing and pick layer wired back into the conversation.

Accomplishments that we're proud of

  • A genuinely conversational interface to REAL physics solvers and the QGIS Processing toolbox -- not a wrapper or a mock.
  • 175+ tools and five-plus hazard families wired end to end: coastal flood, urban stormwater, groundwater, seismic, and tsunami.
  • A scale-to-zero AWS architecture with near-zero idle cost that scales out only on demand.
  • On-map vector drawing, a code-gated and per-session-isolated public demo, and live map rendering with legends and time animation.

What we learned

Learning how to fit the user's prompt with the best model for the job required extensive research of each model's capabilities and limitations. We also learned that the hard part of an AI modeling tool isn't the model it's the data and infrastructure plumbing: making real solvers reproducible, scalable to zero, and reliable enough to trust. Strong tool contracts and honest degradation matter more than clever prompting.

What's next for TRID3NT

  • Move the QGIS Processing toolbox onto AWS Batch (Spot) -- the same scale-to-zero substrate as the solvers -- so heavy and plugin algorithms (a guided digitizer, an AI canopy detector) run off the downscaled agent box.
  • More engines: HEC-RAS / HEC-HMS, TELEMAC, Landlab, HYSPLIT.
  • Compound, multi-driver scenarios (surge + river + rainfall together).
  • Per-user agent isolation (ephemeral compute per session) for real concurrency.
  • Exportable, defensible reports and 3D visualization.

Built With

Share this project:

Updates