FloatChat — Ocean data, answered.

Problem. ARGO floats generate rich ocean data, but it lives in specialist NetCDF files with QC rules and jargon. Non-experts (and even busy experts) rely on screenshots or outdated charts—fueling mistakes and misinformation.

Our idea. FloatChat is an agentic AI dashboard that turns plain-English questions into validated, cited answers with maps, charts, and reproducible CSV/NetCDF exports. Every result shows what data, from where, and how fresh—so insight is fast and trustworthy.

What the user can do.

• “Show salinity profiles near the equator in March 2023.”
• “Nearest ARGO floats to Mumbai today.”
• “Compare temperature & salinity trends in the Arabian Sea over 6 months.”
• "One-click export to CSV/NetCDF; jump from chat to map and 3D globe."

How it works (agentic pipeline).

1. Query Planner (LangGraph) clarifies variables, region, and time.
2. Schema RAG maps intent to our SQL schema (tables/columns).
3. SQL Validator (sqlglot rules, allow-lists, unit checks) prevents errors.
4. Executor runs on PostgreSQL + PostGIS (data from NetCDF → SQL/Parquet).
5. Provenance Agent adds dataset version, QC status, freshness, and citations (Argo/ERDDAP/Argovis, Copernicus fallback).
6. Explainer returns a short summary + chart (Plotly) + download.

Tech stack.

Data/ETL: Python, xarray, netCDF4, Pandas → PostgreSQL/PostGIS, Parquet.
Retrieval: pgvector/Chroma for metadata & docs.
Agents/LLM: LangGraph + Azure OpenAI (GPT-4 + embeddings).
UI: Next.js, Tailwind, shadcn/ui, MapLibre/Mapbox GL, Plotly.
Deploy: Azure Container Apps + managed Postgres; MCP connectors to add gliders/buoys/satellite later.

Why it fits the track. The system plans, validates, and cites every answer, surfaces uncertainty & QC, and can cross-check with Copernicus—reducing cherry-picking and unit/time mistakes.

Impact. Earlier hazard signals for disaster teams, fuel/time savings via better routing, faster fisheries decisions, and tighter spill/SAR modeling—delivered in minutes, not hours.

PoC scope for hackathon. Indian Ocean ARGO subset, three exemplar tasks (profiles, nearest floats, trend compare), live map + 3D globe, exports, and a 5-minute demo showing NL→SQL with citations.

Success metrics. Time-to-answer (sec), % answers with citations, data freshness, validation pass rate, and user repeat queries.