dmj.one

Work / Bharat-First / JalNetra

No. 09 · Bharat-First · Edge AI · AMD Slingshot 2026

The well sees what the lab cannot. जलनेत्र · JalNetra

Rural India is tested for water quality two or three times a year. By the time the disease shows up, the lab is already too far away. JalNetra puts an AMD Ryzen AI NPU and a handful of cheap sensors at the borewell. Inference under 100 milliseconds. Continuous. Solar-viable. Offline by default.

View source AMD Slingshot 2026 How it works
SENSOR 01 ESP32-S3 · pH · TDS SENSOR 02 turbidity · temp SENSOR 03 level · flow LoRa 866 MHz long range · low power EDGE GATEWAY AMD Ryzen AI XDNA NPU · ONNX RT FastAPI · SQLite < 100 ms inference 5·15 W power GCP · BigQuery periodic sync only FARMER SMS · WhatsApp voice · 22 langs

Sensor → LoRa → Edge NPU → Farmer. Cloud is optional.

<100 msInference at the well
5·15 WPower · solar-viable
₹70KPer village · 5 sources
30·40%Water saved on irrigation
22Indian languages for alerts

Act I · The Lab Is Too Far

By the time the test result comes back,
the disease is already in the village.

  1. 2·3

    Times per year a typical rural water source is tested today.The protocol exists. The labs are far. The samples spoil. The reports return after the season.

  2. 24/7

    Frequency at which contamination, depletion and salinity actually change.Continuous sources need continuous monitoring. Quarterly testing is theatre.

  3. 5

    Water sources per village JalNetra covers for a single deployment.Borewells, hand-pumps, canals, reservoir outlets, irrigation channels. Five hundred to two thousand people served.

  4. 0

    Bars of mobile signal at most rural sources. The cloud is not an option.Alerts must work over LoRa, SMS, WhatsApp and voice. Inference must run where the water flows.

India does not need a fancier dashboard in Delhi. It needs a sensor at the handpump in Sirmaur that talks to the farmer in Pahari, before the child gets sick.

Act II · The Edge

Three small models.
One cheap NPU. Zero cloud dependency.

JalNetra runs every model on-device through ONNX Runtime on the AMD XDNA NPU. The cloud is a backup, not a requirement. The same gateway answers a question, files an alert, optimises tomorrow's irrigation cycle.

Model 01

Anomaly Detector

Flags abnormal pH, TDS, turbidity, temperature in real time. Pattern-matches months of normal behaviour at this exact source.

Isolation Forest / Autoencoder
Model 02

Depletion Predictor

Forecasts groundwater level trends days ahead, so the panchayat can schedule rotational supply before the well runs dry.

XGBoost regression
Model 03

Irrigation Optimiser

Recommends when, how much and which channel to open. Field trials show 30 to 40 percent water savings without yield loss.

Multi-objective optimisation

Act III · A Reading In Motion

From a drop in the borewell
to a voice in the farmer's ear.

  1. T = 0 s · the source

    The sensor takes a reading

    An ESP32-S3 with pH, TDS, turbidity, level and temperature probes samples the source. The Heltec WiFi LoRa 32 V3 board transmits the packet over LoRa 866 MHz.

  2. T < 0.1 s · the gateway

    The edge NPU decides

    The gateway runs all three ONNX models against the new reading. The anomaly score, depletion forecast and irrigation recommendation are computed in under 100 milliseconds at 5 to 15 watts.

  3. T < 1 s · the dashboard

    The local React dashboard updates

    A WebSocket stream pushes the new reading and any alerts to the panchayat dashboard. SensorCard, MapView and TrendChart components render on the local network without contacting the cloud.

  4. T < 2 s · the farmer

    The alert reaches the farmer

    If the model flags contamination or depletion, the gateway dispatches an alert through the Bhashini API in the farmer's mother tongue, over SMS, WhatsApp and voice. Twenty-two Indian languages.

  5. when connectivity returns

    The cloud catches up later

    Periodic sync to GCP BigQuery, Cloud Functions and Cloud Storage gives the district officer aggregated trend data. The cloud is a reporting layer, not the critical path.

Act IV · The Stack

Open source. Field-serviceable.
Reproducible Makefile.

  • AMD Ryzen AI · XDNA NPU
  • ONNX Runtime
  • XGBoost · scikit-learn
  • Python 3.11+
  • FastAPI · async
  • aiosqlite
  • React 19 · TypeScript
  • Vite · Tailwind
  • Recharts · Leaflet · Zustand
  • ESP32-S3 · Heltec WiFi LoRa 32 V3
  • PlatformIO · Arduino
  • LoRa 866 MHz · SX1262
  • Docker · Docker Compose
  • Nginx · Certbot · systemd
  • Terraform · GCP
  • BigQuery · Cloud Functions · GCS
  • Bhashini API · 22 languages

Act V · Proof

Built for the AMD Slingshot.
Designed for the village.

AMD Ryzen AI Slingshot 2026

Built for the AMD Slingshot 2026 hackathon under the Sustainable AI and Green Tech theme. Edge inference is the entire premise, not a feature.

Open API surface

FastAPI exposes /api/v1/readings, /alerts, /nodes, /predictions, /reports, /sync, plus a WebSocket /ws/live for real-time sensor streams. Swagger and ReDoc documentation served from the gateway itself.

One Makefile to rule the deployment

make install, make dev, make run, make train, make docker-up, make deploy-gcp, make deploy-terraform. Field engineers do not need a five-page runbook to bring a village online.

Built by team dmj.one

Divya Mohan and Kumkum Thakur. Idea brief, system design document, eligibility, themes, evaluation criteria and a JalNetra-specific design doc all live in /docs of the public repository.

If a borewell can speak Pahari, your factory floor can speak anything.

I build edge-AI products that put the model where the signal lives. Sub-100 ms decisions, sub-15 W power, offline-first, multilingual by design. The same pattern works for a borewell, a substation or an OEM line.

Hire me Back to work