In DevelopmentApril 5, 2026

Float-AI

Plain-English questions over the ARGO ocean fleet. Every answer ships with the SQL it ran — RAG router (LangChain + Gemini + FAISS), hardened text-to-SQL, NetCDF→Postgres ETL.

FastAPIPostgreSQLReactLeaflet

Source

Yashwanth Kamireddi · ML Engineer · Full-Stack Developer

Overview

Float-AI turns plain-English questions into audited SQL, maps, time-series, and float profiles over the ARGO oceanographic fleet — every data answer ships with the exact SQL it ran, for full provenance. Built for SIH 2025 and qualified to Round 2 of India's national hackathon.

The problem

The ARGO programme runs the world's largest ocean-observing array — 4,000+ autonomous floats producing temperature and salinity profiles every 10 days. The data is open and rich; the access path is gated by SQL fluency, dataset shape, and tooling. Float-AI replaces that path with conversation.

How it works

LLM + RAG pipeline

A LangChain router classifies each turn as conversational or data intent. Data intents are routed to a generation step grounded by a FAISS index over schema documentation and domain notes — so the model produces SQL informed by the actual columns and their meanings, not vibes.

Hardened text-to-SQL

SQL generation is constrained to a fixed column allowlist (float_id, profile_date, latitude, longitude, pressure, temperature, salinity) — the model literally cannot reference arbitrary columns. Every answer carries the SQL it ran, as a receipt.

Real ETL

A COPY-based pipeline ingests NetCDF ARGO profiles directly into PostgreSQL — fast bulk loads with optional truncate/append modes for fast rebuilds. Mandatory fields are mapped and coerced (date, lat/lon, pressure), with optional temperature/salinity.

Map + plots, not just text

The frontend is a Vite + React SPA with a Leaflet ocean map (live fleet markers, trajectories, health filters), Plotly depth-aware profile and time-series charts, and a Cmd-K command palette. A health endpoint exposes DB / FAISS / API-key status before the chat loads.

Architecture

Backend — FastAPI serving /api/ask plus operational routes for stats, floats, profiles, trajectories, quality, and time-series
AI — Gemini via langchain-google-genai; FAISS vector store; HuggingFace embeddings; LangChain router for intent classification
Data — PostgreSQL argo_profiles table; SQLAlchemy access; ETL via data_pipeline/build_database.py
Resilience — Sample-data fallback for API routes when the DB is unavailable; the UX stays responsive even with the database offline

Recognition

Smart India Hackathon 2025 — Round 2 Qualifier with this build (Float Chat).

Back to all projects