Candor 🧠📱

Candor is a privacy-first Android recovery companion built for on-device reflection, pattern tracking, and insight generation. Users log a daily check-in, review their recent history, and generate AI-assisted observations — all without sending their data to the cloud.

This project is based on Google AI Edge Gallery and has been refined into a focused recovery experience for a hackathon setting. The core idea is simple:

Log your day → detect patterns → surface insights — all on-device.

---

✨ What Candor Does

• Guides users through a daily recovery check-in with progressive disclosure.
• Tracks five core signals:
  • Craving intensity
  • Mood
  • Stress level
  • Social connection
  • Self-efficacy
• Supports trigger tagging alongside structured inputs.
• Captures freeform reflection for added context.
• Stores all entries locally on device 🔒
• Generates:
  • Streaks and consistency summaries
  • Early warning signals
  • AI-powered insights using an on-device Gemma model
• Allows users to review and edit past entries

---

🔐 Why It Matters

Most recovery tools force a tradeoff between privacy and usefulness. Candor is designed to eliminate that:

• All personal data stays on-device
• AI insights are generated locally
• No need to trust a cloud-based pipeline with sensitive information

The only network dependency is during initial model download via Hugging Face.

---

📲 Product Walkthrough

1. Log

The Log tab is the primary entry point.

• Starts with a simple “Log Your Day” CTA
• Reveals questions progressively (low friction)
• Uses intuitive sliders for core metrics
• Supports structured trigger tagging
• Saves one entry per day (updates if re-logged)

---

2. Insights

The Insights tab combines rule-based logic with LLM reasoning.

• Requires ≥ 5 check-ins before AI insights activate
• Builds prompts from up to 30 recent entries
• Uses hardware acceleration in this order:
  NPU/TPU → GPU → CPU ⚡
• Stores generated insights locally for persistence

Model Used:

• Gemma-4-E2B-it
• gemma-4-E2B-it_qualcomm_sm8750.litertlm

---

3. History

The History tab lets users:

• Scroll through past entries
• Edit previous days
• View streaks and consistency trends 📊

---

⚙️ Technical Highlights

• Kotlin + Jetpack Compose Android app
• Hilt for dependency injection
• Proto DataStore for check-in persistence
• Room database for insight storage
• LiteRT / Google AI Edge runtime for on-device inference
• Gemma 4 E2B for local reasoning

---

🏗️ Architecture At A Glance

Candor is structured around a focused recovery flow:

• RecoveryApp.kt
  → Main 3-tab experience (Log, Insights, History)

• RecoveryViewModel.kt
  → Manages check-in state and saving logic

• InsightsViewModel.kt
  → Handles model loading, prompt building, and insight generation

• InsightsEngine.kt
  → Rule-based pattern detection

• CheckInRepository.kt
  → Local check-in storage via Proto DataStore

• RecoveryDatabase.kt + InsightRepository.kt
  → Room storage for generated insights

For more detail, see:
👉 docs/ARCHITECTURE.md

---

📁 Repository Guide

• DEVELOPMENT.md → Setup + build instructions
• docs/ARCHITECTURE.md → System design + data flow
• docs/IMPLEMENTATION_HISTORY.md → Development evolution
• Android/README.md → Android-specific notes
• Bug_Reporting_Guide.md → Debugging guidance
• Function_Calling_Guide.md → AI Edge Gallery notes

---

🚀 Quick Start

Requirements

• macOS or Linux
• Android Studio
• Android SDK (API 35)
• Android 12+ device or emulator
• Hugging Face OAuth app

---

Clone

git clone https://github.com/akhilkadari/Candor.git
cd Candor

Build

From the Android project root:

cd Android/src
./gradlew assembleRelease

For detailed setup, including OAuth placeholders, see DEVELOPMENT.md.

Demo Notes For Reviewers

If you are reviewing Candor for a hackathon submission, the intended flow is:

1. Open Candor and start a daily check-in from the Log tab.
2. Save several entries or seed mock data during development.
3. Load the preferred Gemma 4 E2B model through the model management flow.
4. Generate insights from the Insights tab.
5. Review history and edit a saved day from the History tab.

Current Status

Candor is in late-stage hackathon polish. The core recovery flow, local persistence, and on-device insight generation are implemented. Current documentation is aimed at making the project easy to evaluate, demo, and extend.

Acknowledgements

• Built on top of Google AI Edge Gallery.
• Uses Google AI Edge / LiteRT infrastructure for on-device inference.
• Uses Gemma-family models for private local analysis.

License

Candor is licensed under the Apache License 2.0. See LICENSE.