IoT & Hardware2026completed

Benthic Guardian

Benthic Guardian is Team Terronix’s SLIoT Challenge 2026 finalist system — an intelligent IoT platform to protect Sri Lanka’s marine biodiversity through real-time coral-reef monitoring and early warning. Rising sea temperatures and bleaching outpace slow, expensive survey methods; Benthic Guardian closes that gap with underwater temperature sensing, live cloud telemetry, a monitoring dashboard, and AI forecasting for heat stress.

Hardware & ingest: ESP32-based sensor nodes POST readings (`sensor_uid` + temperature) to a FastAPI backend (`POST /data`). Approved devices write into PostgreSQL (Supabase in production; SQLite locally). Auth, sensor provisioning, scheduled prediction/alert jobs, and CORS-ready deploy on Render complete the API layer used by both field devices and the web app.

Dashboard: a Next.js + TypeScript UI for login, live readings, maps, analytics, forecasts, and alerts — wired to the backend via `NEXT_PUBLIC_API_URL`.

AI / digital twin (model repo): satellite SST + DHW stand in for seabed loggers in the prototype. A Physics-Informed Neural Network (PINN) builds a continuous space–time temperature field (advection–diffusion residual) to fill between sensors; an ANN–LSTM with 60-day history forecasts SST/DHW at +1 / +3 / +7 days (beating persistence at 3–7 days in evaluation); a NOAA-style DHW risk scorer maps Safe / Watch / Alert. Sustainable deployment research includes microbial fuel cells, solar power, and marine-grade cabling toward long-term reef use.

IoTESP32FastAPIPINNNext.jsSLIoT 2026
2018

Year

0

Features

0

Technologies

Problem
  • Detecting bleaching risk early enough for action, not only after reefs are already damaged
  • Building a credible digital twin when real seabed loggers are still limited — satellite SST/DHW as proxy
  • PINN alone underperforms short-term forecasts vs persistence; map and forecast roles must stay separated
  • Field-to-cloud reliability: approved sensors, every-sample ingest, CORS, and cold-start deploys
  • Scoring health with sustained heat (DHW), not absolute temperature alone
Solution
  • End-to-end IoT pipeline: ESP32 → FastAPI → Supabase → dashboard + scheduled ML jobs
  • Triangle + noise sensor simulation from SST for prototype; path to replace with live loggers
  • PINN for spatial fill/maps; 60-day ANN–LSTM for +1/+3/+7 forecasts that beat persistence at 3–7 days
  • Device POST without JWT keyed by registered sensor_uid; provision scripts for field prototypes
  • NOAA-style DHW Watch/Alert thresholds combined with anomaly, duration, and warming rate

Key Features

ESP32 underwater temperature sensing with live cloud transmission

FastAPI ingest (`POST /data`), auth, sensor provisioning, and scheduled prediction/alert jobs

PostgreSQL / Supabase storage with migration tooling from local SQLite

Next.js dashboard: login, sensor readings, maps, analytics, forecasts, and alerts

PINN spatial twin — continuous (lat, lon, time) → temperature with heat PDE residual

ANN–LSTM 60-day forecaster for +1 / +3 / +7 day SST & DHW

DHW-aware bleaching risk scoring (Safe <4, Watch 4–8, Alert >8)

Risk maps / Folium visualization and hold-out evaluation notebooks

Prototype triangle multi-sensor simulation from satellite SST (+ noise) until real loggers

Sustainable hardware research: microbial fuel cells, solar, marine-grade cabling

Deploy path: Render (Docker/TensorFlow backend) + Vercel dashboard + field ESP32 POSTs

Technologies

ESP32C / Arduino (firmware)FastAPIPythonPostgreSQLSupabaseSQLiteSQLAlchemy / psycopgJWT AuthAPSchedulerNext.jsTypeScriptTailwind CSSTensorFlow / KerasPINN (physics-informed NN)ANN–LSTMNumPy / Pandas / SciPyscikit-learnFoliumRenderVercelDocker

Learnings

  • Shipped an end-to-end AIoT stack from ESP32 nodes through cloud API, dashboard, and ML forecasts

  • Combined physics-informed spatial modeling (PINN) with sequence forecasting (ANN–LSTM) for complementary map vs day-ahead tasks

  • Designed device ingest without JWT while keeping sensors registered and approved server-side

  • Operated production-minded data layering: SQLite locally, Supabase Postgres + Render for deploy

  • Evaluated forecast skill against persistence/climatology and tied risk to DHW thresholds, not raw °C alone

  • Worked as Team Terronix across hardware, backend, ML, and marine-conservation context for SLIoT finals

Highlights

SLIoT 2026 Finalist

PINN + ANN–LSTM

Real-time Reef Alerts

Next Project

Kapu — Kapruka AI Shopping Agent

Full-stack conversational shopping companion for Kapruka: live catalog search, multi-agent routing, delivery checks, voice input, and guest checkout — English, Singlish, or Sinhala.