Askari AI — Ranger Intelligence Copilot

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.

---

Demo

Video walkthrough

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

---

Screenshots

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

---

Features

Feature	Stack
Natural language patrol queries	Apple Foundation Models + tool calling (local SQLite)
Voice incident note dictation (STT)	AVAudioRecorder + SpeechAnalyzer / SpeechTranscriber
Pre-patrol AI briefing	Apple Foundation Models + ephemeral session
Offline-first sync	PowerSync Sync Streams → Supabase
Reactive UI	SwiftUI + The Composable Architecture

---

Project Structure

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)

---

Getting Started

Prerequisites

• 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

---

1. Clone and open in Xcode

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.

---

2. Configure Secrets

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

---

3. Set up Supabase

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

File	Purpose
supabase/schema.dbml	Schema documentation — tables, PKs, FKs, checks (DBML format)
supabase/schema.sql	CREATE TABLE statements + spot_types seed data
supabase/indexes.sql	Performance indexes
supabase/functions.sql	Helper functions (auth_staff_id, auth_staff_park_id, auth_staff_rank, trigger_set_updated_at, handle_new_user)
supabase/triggers.sql	updated_at triggers + commented on_auth_user_created auth trigger
supabase/rls.sql	RLS enables + all policies
supabase/powersync.sql	PowerSync scoping notes (not executed in DB)

3a. Run the setup SQL in order

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.

Step	File	Action
1	supabase/schema.sql	Creates all tables; seeds 14 spot_types rows
2	supabase/indexes.sql	Adds performance indexes
3	supabase/functions.sql	Creates helper functions + handle_new_user
4	supabase/triggers.sql	Attaches updated_at triggers; see note for auth trigger
5	supabase/rls.sql	Enables RLS and applies all policies

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.

3b. Create a Storage bucket (optional — for media attachments)

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

3c. Set up your first admin/ranger

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';

---

4. Set up PowerSync

4a. Link your Supabase project

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

4b. Deploy the sync stream config

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.

---

5. Build and run

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

---

Swift Package Dependencies (Xcode SPM)

Package	URL	Version
PowerSync Swift SDK	https://github.com/powersync-ja/powersync-swift	1.13.0
Supabase Swift	https://github.com/supabase/supabase-swift	>= 2.41.1
TCA	https://github.com/pointfreeco/swift-composable-architecture	>= 1.25.1

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+.

---

Backend

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

---

PowerSync Sync Streams

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.

---

Apple Intelligence — On-Device AI

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

Foundation Models (LLM)

Feature	Session type	Notes
Ranger Copilot chat	LanguageModelSession (persistent)	Multi-turn; retains transcript
Pre-patrol briefing	LanguageModelSession (ephemeral)	Fresh session per briefing
Dashboard queries	LanguageModelSession (ephemeral)	Fresh session per query

• 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

SpeechAnalyzer (STT)

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

Component	Role
AVAudioRecorder	Captures 16 kHz mono PCM WAV to a temp file
SpeechAnalyzer	Orchestrates transcription pipeline
SpeechTranscriber	On-device locale model; options-based init with .volatileResults
AssetInventory	Downloads locale model assets on first use (~50–100 MB); no-op if already current
STTBufferConverter	Resamples AVAudioPCMBuffer for live streaming; primeMethod = .none avoids timestamp drift

Voice Note Dictation Flow

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.