Software platform

From farmer intent to verified field action

One workspace holds server, website, app and device modules, each with its own README. No module invents its own representation of device IDs, commands, events or states — shared schemas are the contract.

09 · Mobile App + Web PlatformPLANNED

Farmer operations and fleet administration

Mobile-first farmer interface plus a web dashboard for administration, support and analytics.

Mobile — P0
  • Farm and pond creation
  • Gateway provisioning and secondary pairing
  • Current known state of devices
  • Safe ON/OFF command for an aerator/motor group
  • Schedules and alerts
Web — P1
  • Administration and farm management
  • Device/fleet management and support tooling
  • Analytics and execution history
  • Audit trail for automation and manual control
/server

APIs, auth, device services, telemetry, commands, automation, AI orchestration

/web

Administration and farm-management dashboard

/mobile

Farmer-facing mobile application

/device

Gateway and secondary-controller firmware

/shared

Protocol schemas, types, validation and common constants

/docs

Architecture, deployment, electrical integration, LoRa protocol, operations

10 · Backend & CloudPLANNED

Device registry, telemetry, commands and audit

APIs, authentication, device services, telemetry, commands, automation and AI orchestration.

EntityKey data
FarmOwner, location, configuration, timezone, status
PondFarm, pond ID, metadata, connected devices, operating profile
DeviceDevice ID, type, firmware, credentials, gateway, last seen, health
Motor / AeratorPond, controller, control channel, operating state, schedules
CommandTarget, requested state, issuer, timestamp, sequence, acknowledgement
EventType, source, timestamp, payload, severity, correlation ID
AutomationTrigger, conditions, action, safety constraints, execution history
UserIdentity, role, farm access, permissions
11 · AI Agent OrchestrationPLANNED

Ordered tasks, shared context, deterministic safety

AI agents must follow the correct order, preserve context, work concurrently where appropriate, and execute accurately rather than behaving as isolated chat sessions.

Orchestration components
  • 01
    Planner
    Decomposes a high-level request into explicit tasks.
  • 02
    Context service
    Supplies relevant farm, pond, device and historical state.
  • 03
    Task scheduler
    Determines dependencies and which tasks may execute concurrently.
  • 04
    Workers / agents
    Perform bounded tasks with typed inputs and outputs.
  • 05
    Safety / policy layer
    Validates commands before field devices receive them.
  • 06
    State manager
    Records task status, retries, failures and completion.
  • 07
    Audit layer
    Records who/what initiated an action and what happened.
Execution states
PLANNEDREADYRUNNINGWAITING / RETRYINGSUCCEEDED / FAILEDVERIFIED

Concurrency is dependency-aware: checking several ponds can run concurrently, while a motor command waits for relevant context and safety checks. Conflicting commands targeting the same actuator require deterministic arbitration.

Context discipline
  • Every task receives a task ID and correlation ID
  • Context is scoped to farm/pond/device, not the whole conversation
  • Tool results affecting future execution are persisted
  • Critical actions use structured parameters, not free-form text
  • Completion is based on verified device acknowledgement where possible
12 · Safety & Offline OperationCURRENT PROTOTYPE

AI and cloud request actions — the field layer decides

Aquaculture field systems operate in wet, electrically noisy and connectivity-variable environments. Reliability must be designed in.

Power restart

Gateway/node recovers automatically to a known safe state.

Internet loss

Local control and essential schedules continue where configured.

LoRa loss

Secondary follows explicit communication-loss policy; no unsafe uncontrolled switching.

Duplicate command

Command ID/sequence prevents repeated execution.

Stale command

Expiry/deadline prevents old commands from being applied later.

Firmware fault

Watchdog and rollback/recovery strategy.

Electrical fault

Existing industrial protection remains authoritative.

Auditability

Every remote control action is recorded with source and result.

AI failure

AI cannot bypass deterministic safety and authorization layers.

CURRENT PROTOTYPE

Critical control principle: AI and cloud services should request actions; the field control layer decides whether an action is currently safe and valid. Identity, authorization, command validity and physical safety are separate controls.