PowerSync AI Hackathon 2026 · Local-First & Offline-Capable AI

An on-device AI assistant for wildlife park rangers. All AI inference runs locally using Apple Foundation Models and SpeechAnalyzer — no connectivity required for intelligence features. Data syncs to Supabase via PowerSync when back in range.

https://github.com/Clemo97/askari-ai/raw/main/Videos/RPReplay_Final1776686584.MP4

Patrol map — live incident pins with colour-coded severity, park boundary overlay, and tap-to-view incident detail

Incident report — full detail view with category, GPS coordinates, field notes, and photo evidence

AI Intelligence dashboard — natural-language query powered by Apple Foundation Models with tool-called results from the local PowerSync SQLite DB

AskariAI/
├── App/                    # App entry point, root navigation
├── Managers/               # SystemManager, SupabaseConnector, Schema
│   ├── Schema.swift        # PowerSync local SQLite schema
│   ├── SystemManager.swift # PowerSync database + sync stream subscriptions + TCA dependency
│   ├── SupabaseConnector.swift
│   └── _Secrets.swift      # ← gitignored, copy from template
├── Features/               # TCA Reducers + Views (co-located)
│   ├── App/                # AppFeature, MainFeature
│   ├── Auth/               # AuthFeature
│   ├── Missions/           # MissionsFeature
│   ├── ActiveMission/      # ActiveMissionFeature (GPS tracking)
│   ├── Dashboard/          # DashboardFeature (AI query + intel overview)
│   └── Incidents/          # LogIncidentFeature (voice dictation + media)
├── Models/                 # Codable data models
├── Views/                  # SwiftUI views
│   └── Copilot/            # CopilotChatView, AIBriefingView
└── AI/
    ├── AIManager.swift     # Foundation Models sessions + SpeechAnalyzer lifecycle
    └── Tools/
        └── CactusTools.swift  # Tool protocol implementations (DB query + incident logging)

supabase/
└── migrations/
    ├── 001_initial_schema.sql   # Applied ✓
    └── 002_rls_policies.sql     # RLS policies

powersync/
└── sync-config.yaml  # Sync Streams config (edition 3)

• Xcode 26+ (Swift 6 toolchain)
• iOS 26+ device — Apple Foundation Models and SpeechAnalyzer require a real device with Apple Intelligence enabled (simulator is not supported)
• Apple Intelligence enabled on the device (Settings → Apple Intelligence & Siri)
• A Supabase project
• A PowerSync Cloud instance linked to that Supabase project

git clone https://github.com/Clemo97/askari-ai.git
cd askari-ai
open AskariAI.xcodeproj   # or .xcworkspace if present

Swift Package dependencies resolve automatically on first open. If they don't, go to File → Packages → Resolve Package Versions.

The app reads credentials from AskariAI/Managers/_Secrets.swift, which is gitignored. A template is provided at AskariAI/Managers/_Secrets.template.swift.

Copy the template:

cp AskariAI/Managers/_Secrets.template.swift AskariAI/Managers/_Secrets.swift

Then fill in _Secrets.swift:

enum Secrets {
    static let supabaseURL      = URL(string: "https://<your-project-ref>.supabase.co")!
    static let supabaseAnonKey  = "<your-supabase-anon-key>"
    static let powerSyncEndpoint = "https://<your-instance-id>.powersync.journeyapps.com"
    // Optional — set to your Supabase Storage bucket name to enable media attachments
    static let supabaseStorageBucket: String? = "incident-media"
}

You can find these values at:

• Supabase URL + anon key: Supabase Dashboard → Project Settings → API
• PowerSync endpoint: PowerSync Dashboard → Your Instance → Connection URL

The full database definition is split into focused files under supabase/:

Open the Supabase SQL Editor (Dashboard → SQL Editor → New query) and run each file in the order listed. Each query window is a fresh paste — run them one at a time.

Auth trigger (step 4): triggers.sql includes a commented-out CREATE TRIGGER on_auth_user_created on the auth.users table. Uncomment and run it in the Supabase SQL Editor (which executes as superuser). This enables automatic staff row creation on signup.

1. Supabase Dashboard → Storage → Create bucket
2. Name it incident-media (or whatever you set in _Secrets.supabaseStorageBucket)
3. Set it to Private
4. Add a storage policy allowing authenticated users to upload/download their own files

Sign up via the app — handle_new_user automatically creates a staff row linked to the first park in the database. To promote an account to admin:

UPDATE public.staff SET rank = 'admin' WHERE email = 'you@example.com';

To assign a ranger to a specific park (if multiple parks exist):

UPDATE public.staff
SET park_id = (SELECT id FROM public.parks WHERE name = 'Nairobi National Park')
WHERE email = 'ranger@example.com';

Follow the PowerSync Supabase integration guide to connect your PowerSync instance to your Supabase database.

npm install -g @powersync/cli
powersync login
powersync link cloud --project-id=<your-project-id>
powersync deploy sync-config

The config is at powersync/sync-config.yaml.

1. Select your target device in Xcode (real device recommended for AI features)
2. Product → Run (⌘R)
3. Sign up with an email — handle_new_user creates a staff row and assigns the default park automatically
4. To promote yourself to admin, run UPDATE public.staff SET rank = 'admin' WHERE email = 'you@example.com'; in the SQL Editor
5. Apple Foundation Models are built into the OS — no model download required for LLM features
6. SpeechAnalyzer locale assets (~50–100 MB) download automatically on first dictation tap

All packages are declared in Package.swift and resolve automatically in Xcode.

Note: FoundationModels and Speech (SpeechAnalyzer) are Apple system frameworks — no SPM dependency required. They are available on iOS 26+.

• Supabase Project: (set supabaseURL and supabaseAnonKey in _Secrets.swift)
• PowerSync Instance: (set powerSyncEndpoint in _Secrets.swift)
• Schema: see supabase/schema.dbml

Migrated from sync rules (bucket_definitions) to a single consolidated sync stream (migrated_to_streams). CTE-style with: parameters resolve per-user via auth.user_id() — ranger and admin data are filtered server-side, no client-side parameters required.

config:
  edition: 3

streams:
  migrated_to_streams:
    auto_subscribe: true
    with:
      ranger_own_features_param: >-
        SELECT staff.id AS staff_id FROM staff
        WHERE staff.user_id = auth.user_id()
          AND staff.rank = 'ranger' AND staff.is_active = TRUE
      ranger_park_data_param: >-
        SELECT staff.park_id AS park_id FROM staff
        WHERE staff.user_id = auth.user_id()
          AND staff.rank = 'ranger' AND staff.is_active = TRUE
      admin_map_features_param: >-
        SELECT staff.park_id AS park_id FROM staff
        WHERE staff.user_id = auth.user_id()
          AND staff.rank = 'admin' AND staff.is_active = TRUE
    queries:
      # Ranger: own incidents
      - "SELECT map_features.* FROM map_features, ranger_own_features_param AS bucket WHERE map_features.created_by = bucket.staff_id"
      # Ranger: park context
      - "SELECT parks.* FROM parks, ranger_park_data_param AS bucket WHERE parks.id = bucket.park_id"
      - "SELECT park_boundaries.* FROM park_boundaries, ranger_park_data_param AS bucket WHERE park_boundaries.park_id = bucket.park_id"
      - "SELECT staff.* FROM staff, ranger_park_data_param AS bucket WHERE staff.park_id = bucket.park_id AND staff.is_active = TRUE"
      # Admin: all park map features + context
      - "SELECT map_features.* FROM map_features, admin_map_features_param AS bucket WHERE map_features.park_id = bucket.park_id"
      - "SELECT parks.* FROM parks, admin_map_features_param AS bucket WHERE parks.id = bucket.park_id"
      - "SELECT park_boundaries.* FROM park_boundaries, admin_map_features_param AS bucket WHERE park_boundaries.park_id = bucket.park_id"
      - "SELECT staff.* FROM staff, admin_map_features_param AS bucket WHERE staff.park_id = bucket.park_id AND staff.is_active = TRUE"
      # Global reference (all users)
      - SELECT * FROM spot_types WHERE spot_types.is_active = TRUE
      - SELECT * FROM mission_role_types

On the client, SystemManager.connect() subscribes to the single stream and awaits its first sync:

let sub = try await db.syncStream(name: "migrated_to_streams", params: nil).subscribe()
try await sub.waitForFirstSync()

auto_subscribe: true means the server begins syncing immediately on connection — the explicit subscribe call is used only to obtain a handle for waitForFirstSync(), which blocks until the local SQLite database has received its initial data.

All AI features use Apple system frameworks — no third-party model hosting, no network calls for inference.

• Availability checked at runtime via SystemLanguageModel.default.isAvailable
• Tool calling via the Tool protocol: QueryRecentIncidentsTool, GetRangerStatsTool, LogIncidentTool — each executes a PowerSync SQLite query and returns structured JSON to the model
• No model download required — Apple Intelligence is part of the OS

Voice note dictation uses SpeechAnalyzer + SpeechTranscriber (iOS 26 Speech framework), following the WWDC25 session 277 reference architecture.

1. User taps Dictate → loadSTTIfNeeded() requests mic permission then ensures locale assets are installed via AssetInventory
2. AVAudioRecorder records to a 16 kHz mono PCM .wav temp file (audio session: .playAndRecord / .spokenAudio)
3. User taps Stop → recorder stops, audio session deactivated
4. SpeechAnalyzer.start(inputAudioFile:finishAfterFile:true) processes the file autonomously; finalizeAndFinishThroughEndOfInput() flushes results
5. Only isFinal results are collected; transcription appended to the incident notes field; temp file deleted

Device requirement: Apple Foundation Models and SpeechAnalyzer require a physical device running iOS 26 with Apple Intelligence enabled. Neither works in Simulator.