This is a submission for the Gemma 4 Challenge

RemoraAI: The Subconscious Social Network Powered by Edge AI

What I Built

Dreams are our most private thoughts.

Yet most AI-powered journaling apps require users to upload deeply personal emotions, fears, and subconscious experiences directly to the cloud.

Remora was built to challenge that assumption.

Remora is a privacy-first “Subconscious Social Network” powered by Gemma 4 running directly on-device using LiteRT-LM and Flutter.

The app allows users to:

• record dreams via voice,
• receive AI-powered psychological interpretation,
• detect recurring subconscious patterns over time,
• generate surreal dream visuals,
• and optionally publish anonymized dreams to a public community feed.

The key innovation is that the sensitive psychological analysis happens entirely on-device.

No raw dream data needs to leave the smartphone.

The Core Problem

Dream journaling has historically remained a private, offline activity because users are understandably uncomfortable uploading vulnerable psychological content to centralized servers.

We wanted to answer a difficult question:

Can modern multimodal AI deliver meaningful emotional analysis while preserving user privacy?

Remora demonstrates that the answer is yes.

Demo

Core Flow

1. User records a dream using voice input
2. Gemma 4 processes the narrative locally
3. The app generates:

• a dream title,
• emotional interpretation,
• thematic tags,
• and subconscious motif detection
  1. User optionally generates AI dream artwork
  2. User may privately store or anonymously publish the dream

Demo Content

• Offline “Privacy Mode”

• AI-generated dream art

• Community feed scrolling

Code

Tech Stack

• Flutter
• LiteRT-LM
• MediaPipe
• Flutter FFI
• FastAPI
• Android AI Core
• Gemini Nano
• Imagen 4
• Vector Embeddings + RAG

Architecture Highlights

Local AI Layer

• Gemma 4 E2B via LiteRT-LM
• On-device inference
• NPU acceleration
• Offline-capable “Privacy Mode”

Cloud Layer

• Optional dream image generation
• Anonymous community feed
• Secure transient speech-to-text fallback

Memory Layer

• Vector embeddings for recurring dream motifs
• Retrieval-Augmented Generation (RAG)
• Long-term subconscious pattern analysis

How I Used Gemma 4

We selected the Gemma 4 E2B model because it sits at the ideal intersection of:

• mobile performance,
• low memory footprint,
• multimodal capability,
• and meaningful reasoning quality.

Previous local models were either:

• too large for mobile deployment,
• too slow for real-time inference,
• or incapable of nuanced psychological interpretation.

Gemma 4 E2B solved all three.

Using LiteRT-LM, the model runs directly on-device through Android NPUs or Android AI Core (Gemini Nano where available).

This enables:

• fully offline dream analysis,
• dramatically reduced latency,
• improved privacy,
• and lower infrastructure cost.

Local Inference Pipeline

FlutterGemma.installModel(
  modelType: ModelType.gemma4,
  fileType: ModelFileType.litertlm,
).fromNetwork(
  'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',
);

The Hardest Engineering Problem

One of the biggest challenges was multimodal audio processing.

Although Gemma 4 supports audio understanding conceptually, current LiteRT community weights lack fully fused audio execution graphs for mobile delegates.

Attempting native audio inference produced:

After investigating Google’s AI Edge Gallery implementation, we discovered:

• unsupported audio tensor routing,
• delegate backend limitations,
• and missing Flutter bindings for CPU/NPU graph splitting.

Instead of abandoning voice dreams entirely, we engineered a Secure Hybrid Loop:

1. Audio is transiently transcribed
2. No raw data is persisted
3. Transcription text returns immediately
4. Gemma 4 performs all psychological interpretation locally

This preserved the most sensitive part of the experience entirely on-device.

Subconscious RAG

Remora is not just a dream diary.

Over time, it becomes a semantic memory system for the user’s subconscious.

Dream entities are vectorized using embeddings:

• characters,
• emotions,
• locations,
• recurring symbols,
• and narrative structures.

If a user repeatedly dreams about:

“A woman in a red coat”

…the system detects the recurring motif and surfaces psychological pattern insights over months or years.

This transforms dream logging from passive journaling into longitudinal subconscious analysis.

Dream Visualization

After local interpretation is complete, users can optionally generate dream artwork using Imagen 4.

The backend converts the interpreted dream into a surreal cinematic visual prompt and generates high-resolution dream imagery.

This creates a hybrid architecture:

Community Layer

By default, every dream remains private.

Users may optionally anonymize and publish dreams to the Remora community feed, creating a surreal stream of humanity’s collective subconscious.

Other users can:

• upvote bizarre dreams,
• react to recurring themes,
• or share dreams with therapists or friends.

This transforms deeply personal subconscious experiences into optional social storytelling.

Why Gemma 4 Matters

Before Gemma 4, building an app like Remora was largely impractical.

The model needed to be:

• lightweight enough for smartphones,
• capable of emotional nuance,
• fast enough for real-time interaction,
• and deployable through modern mobile inference stacks.

Gemma 4 E2B made that architecture possible.

It allowed us to move psychological AI away from centralized cloud systems and directly into the user’s pocket.

That shift fundamentally changes what privacy-first AI applications can become.

Future Work

We plan to expand Remora with:

• native multimodal audio execution,
• local image generation,
• lucid dream detection,
• and cross-dream narrative mapping.

As edge AI tooling matures, applications like Remora will increasingly blur the line between local software and personal AI companions.

Final Thoughts

Building Remora with Gemma 4 demonstrated something important:

Edge AI is no longer experimental.

For the first time, mobile devices are capable of delivering meaningful multimodal AI experiences while preserving user privacy by default.

That opens the door to an entirely new generation of personal AI applications.