Selected work / 04

Supraconscious Avatar AI

A multi-surface reflection product that turns journal entries into guided questions while giving the operator a separate place to govern prompts, sources, safety, and review.

My role: Independent product engineer
Scope: Product strategy, AI system design, and full-stack implementation
Stage: Deployed web application / mobile scaffold
Core technologies: Next.js · TypeScript · PostgreSQL + Prisma · OpenAI

Case study

Building a reflection product that can be governed.

The problem

A journaling product can generate polished language and still fail its user if safety, source selection, memory, and prompt changes are hidden inside an opaque model call. Sensitive reflection needs a clear experience for the person writing and a disciplined operating surface behind it.

What I built

I designed and built a multi-app system with a Next.js journal product, a separate admin and content-governance console, an MCP-compatible service, shared AI packages, and PostgreSQL persistence. The active flow uses one consistent guide while source eligibility, safety, prompts, and review remain visible to the operator.

Key decisions

01One guide, one product voice
The active reflection flow uses one consistent guide instead of exposing internal role orchestration as a cast of competing personas.
02Govern retrieval before expanding it
Source review, rights, quote rules, safety intensity, and traceability are enforced before future vector search or broader graph retrieval is introduced.
03Separate the experience from the control plane
The journal stays focused on reflection while a distinct admin application owns prompts, sources, feature flags, quality review, and operational state.

Inside the work / 04 views

Inside the governed reflection product

The authenticated product carries a reflection from private writing into a grounded Guide response, saved history, an explicit seven-dimension model, and user-controlled privacy and security settings.

Reflection history

The private dashboard makes continuity legible without turning reflection into a score: entries, active patterns, the latest Guide response, and saved sessions stay visible together.

Constant Guide

The Guide stays consistent while the seven dimensions remain equal, explicit, and selectively applied to each reflection instead of becoming hidden personas or a ranked progression.

Privacy and preferences

Privacy is a product surface: the account explains processing, supports passkeys, keeps pattern memory opt-in, and exposes the Guide tone and reflection intensity as user-controlled settings.

Technical view

Architecture

Governed reflection platform13 components · 14 connections

Customer surfaces share a policy-first AI package and database, while a separately deployed admin plane controls prompts, sources, safety, and feature state.

Experience surfacesUser and compatible AI clients

Shared AI runtimePolicy before generation

Durable knowledge + stateServer-only persistence

Operator control planeSeparately deployed admin

Next.js
Journal web
Writing, reflection, privacy controls
Express · MCP
ChatGPT surface
Authenticated tools + widget
Flutter · scaffold
Mobile client
Scaffolded product surface
Next.js route handlers
Server routes
Auth, journal, voice, billing
OpenAI · Zod
Safety classifier
Intensity and response boundaries
Shared AI package
Reflection guide
One active voice + structured output
Policy-first keyword RAG
Governed retrieval
Eligible reviewed sources + citations
GraphRAG · feature-flagged
Ontology context
Approved neighborhoods only
Shared AI package
Pattern memory
Traces and reusable reflection state
PostgreSQL · Prisma
Application record
Journal, sources, prompts, traces
PostgreSQL
Reviewed knowledge
Chunks, provenance, ontology state
Separate Next.js app
Admin console
Prompts, sources, safety, quality
Stripe
Subscription boundary
Plans and billing operations

surfaceservicedataaiintegrationcontrolruntime

Read system connections

Journal web → Server routes
ChatGPT surface → Reflection guide
Mobile client → Server routesscaffold
Server routes → Safety classifier
Safety classifier → Reflection guide
Reflection guide ↔ Governed retrieval
Governed retrieval → Reviewed knowledge
Ontology context → Reviewed knowledgeflagged
Reflection guide → Pattern memory
Pattern memory → Application record
Server routes → Application record
Admin console ↔ Application record
Admin console ↔ Reviewed knowledge
Server routes → Subscription boundary

Built so far

Combines a user-facing journal, a separate admin and content-governance application, an MCP-compatible service, shared AI packages, and PostgreSQL persistence.
Gates retrieval by source review, rights metadata, quote permissions, safety intensity, feature state, and citation traceability.
Keeps prompt, source, safety, feature, quality, and subscription operations in a separate admin control plane.

Where it stands

The web product is deployed, while the Flutter mobile client remains a scaffold rather than a released mobile application.
Retrieval is currently policy-first and keyword based; vector search is planned, and graph-assisted runtime context is disabled by default.

Repository

Continue exploring

Next case study

View all selected work

Product strategy / Private services

Argent Matchmaking

Synthetic concept prototype

Open case study