A private voice journal
that stays on your phone.
OffRecord AI Journal transcribes, searches, and reflects on your entries entirely on-device. The source is public on GitHub.
The code.
sourceOffRecord AI Journal is simple at its core: open the app, press record, talk about your day — and Apple’s SpeechAnalyzer transcribes everything on-device using SpeechTranscriber or DictationTranscriber. No speech server. No account. No subscription. Your voice never leaves your phone for transcription.
The source is public, so you can read how it works instead of taking it on trust.
The Friday vision.
north starAfter I started building OffRecord AI Journal, I realized it could go far beyond just recording voice entries.
If you talk to this app every day — about your thoughts, your decisions, how you react to things, what makes you happy, what frustrates you — it builds a picture. A detailed, honest picture of who you are.
That's when I built Friday. A private assistant with a multi-dimensional model of communication style, emotional signature, a personal knowledge graph of people and topics, and temporal mood predictions. All derived on-device from NLTagger signals. All serialized as Codable JSON in a single Core Data entity. All yours.
“I want Friday to become the assistant you can confide in: observant, discreet, and grounded in what you've actually shared. Private, on-device, yours.”
We already leave digital traces everywhere — social media, messages, emails. But all of that is scattered across platforms owned by other companies. None of it is organized. None of it is truly yours.
What if you had one place where you deliberately stored the most honest version of yourself? Not the curated version you put on social media. The real version.
I believe keeping a private, organized record of your inner life will be an important part of the near future. Not science fiction — practical. And OffRecord AI Journal is built to make that happen.
Why iOS only.
platformOffRecord AI Journal is on iPhone and iPad only. There's no Android version — yet.
That's a deliberate choice. Apple gives me strong security guarantees — the Secure Enclave, the Neural Engine, the privacy frameworks. The whole promise of OffRecord AI Journal is that your data never leaves your device. I need to be absolutely sure I can keep that promise before putting it on another platform.
I'd rather be on one platform and keep your data safe than be on two platforms and compromise.
Android will come — after iOS is solid, after the user base grows, and after I'm confident the same privacy guarantees can hold.
What I believe.
principlesPrivacy is the architecture.
Not a feature. Not a toggle. The entire stack is built so your data physically cannot leave your device.
Mental health tools should be free.
A diary shouldn't have a paywall. Nobody should choose between their wallet and their mental health.
AI should work for you, not on you.
AI that processes your thoughts should run on your device, under your control. Not on a server training someone else's model.
Speaking is more honest than writing.
When you write, you edit. When you speak, you're real. The best insights come from unfiltered expression.
The roadmap.
build logVoice journaling + on-device AI
Tech: SpeechAnalyzer with SpeechTranscriber and DictationTranscriber for on-device transcription. NLTagger for sentiment, NER, POS. NSPersistentCloudKitContainer wrapping SQLite. AVFoundation AAC 44.1 kHz capture. CryptoKit AES-256-GCM backups. LocalAuthentication for Face ID / Touch ID. WidgetKit + AppIntents. All inference stays on-device.
Friday + personality model
Tech: Custom FridayAssistantEngine. CommunicationStyle (TTR, expressiveness, directness, formality, signature words). EmotionalSignature (valence / arousal / dominance baselines, time-based patterns, trigger topic correlations). PersonalKnowledgeGraph via NER mapping entities with emotional weights. FridayPredictions for day-of-week mood forecasting. All models as Codable JSON in the Core Data AIState entity.
Talk to Friday + social sharing
Tech: FridayChatView with pattern-matched query system. FridayResponseGenerator assembles natural-language replies from Friday data. ShareablePersonalityCardView renders cards via SwiftUI ImageRenderer at 3x (Instagram Stories 1080×1920, X 1200×675). UIActivityViewController for native share sheet. SKStoreReviewController at milestone entries with 90-day cooldown.
Semantic search + proactive insights
Tech: NLEmbedding for 512-dim sentence embeddings per entry, persisted in Core Data. Custom cosine-similarity vector index for semantic queries. Z-score anomaly detection against emotional baseline triggering local notifications. K-means clustering on embedding space for hidden thematic groupings. Retrieval foundation for v2.0's RAG architecture. Zero network calls.
Multi-language + wrist capture
Tech: Xcode String Catalogs — Hindi, Spanish, Japanese, German first. Speech framework already supports 60+ transcription languages. WatchKit companion via WatchConnectivity. Quick voice entry from wrist. Complications for mood and streak.
Foundation Models + tool calling + SpeechAnalyzer
Tech: Apple's on-device 3B Foundation Model via LanguageModelSession. Friday becomes a real chatbot with multi-turn transcript memory. Autonomous Core Data querying via Tool protocol. Dynamic instructions fed from FridayAssistantEngine for tone matching. @Generable for type-safe structured outputs. streamResponse() for real-time streaming. Requires iPhone 15 Pro+. Zero network calls.
LoRA fine-tuning — Friday learns your tone
Tech: Apple's Foundation Models Adapter Training toolkit for Low-Rank Adaptation. Export 100–1,000 entries as JSONL prompt/response pairs. Train a personal adapter on Mac (32GB+ Apple Silicon). ~160 MB adapter delivered via Background Assets. Loaded via SystemLanguageModel(adapter:). Friday doesn't just know your data — she understands your tone. Training data never leaves your Mac.
A private assistant with memory
End state: an on-device AI assistant grounded in years of your own entries. Full RAG. Personal LoRA adapter. Tool calling for data retrieval. Session transcript with context condensation. Encrypted. On your device. Exportable only by you. Your data never touches a server. Your private assistant archive is yours alone.
Get in touch.
contactI'm Saransh, and I build OffRecord AI Journal.
If you have questions, feedback, or just want to say hi —
Start talking.
Free. Private. No account needed. Just open and talk.
Download free on the App Store →