Diary-to-Empathy AI Pipeline

About This Project

This is my personal implementation of a 13-step modular AI pipeline designed to transform a single diary transcript entry into an emotionally intelligent, concise AI reply along with a dynamically updated long-term user profile. The system addresses two key scenarios:

• Cold start: Processing the first-ever diary entry with no prior user data.
• Mature profile: Processing the 100th diary entry after accumulating rich historical context.

The project emphasizes clean modular design, strict logging as per specifications, and mocking external APIs to ensure zero operational cost and ease of local development.

---

Design Assumptions

Embeddings fixed at 384 dimensions (MiniLM standard)

Empathetic reply length capped at ~55 characters

User profile state accumulates incrementally

Entry IDs are auto-incremented per user starting from entry_1

Logging format strictly adheres to specification for ease of testing

---

Core Pipeline Workflow

The pipeline consists of 13 distinct stages, each with a specific role:

Step	Purpose	Input	Output
RAW_TEXT_IN	Accept raw diary transcript	String (raw text)	Raw transcript string
EMBEDDING	Generate semantic embedding vector (mocked MiniLM)	Raw text	384-dimensional embedding
FETCH_RECENT	Retrieve last 5 parsed entries for context	User ID	Array of recent parsed entries
FETCH_PROFILE	Fetch or initialize user profile	User ID	User profile JSON
META_EXTRACT	Extract lexical and punctuation metadata	Raw text	Metadata JSON
PARSE_ENTRY	Parse transcript into structured JSON with theme, vibe, etc.	Raw text	Parsed entry JSON
CARRY_IN	Evaluate thematic or emotional similarity for context carry-in	Current & recent entries	Boolean carry-in flag
CONTRAST_CHECK	Detect emotional shift vs. dominant profile vibe	Dominant vibe, new vibe	Boolean contrast flag
PROFILE_UPDATE	Update profile counters and traits with new entry data	Parsed entry, existing profile	Updated user profile
SAVE_ENTRY	Persist parsed entry in mock DB with auto-incremented ID	Parsed entry, User ID	Entry ID string
GPT_REPLY	Generate concise empathetic response (≤ 55 characters)	Parsed entry	AI response string
PUBLISH	Package final outputs for downstream use	Entry ID, response, carry-in	JSON output package
COST_LATENCY_LOG	Log mock API cost and latency	—	Log output

Each step emits detailed logs following the pattern:
[TAG] input=<...> | output=<...> | note=<...>

---

Data Structures & Types

ParsedEntry

Semantic interpretation of a diary entry:

interface ParsedEntry {
  theme: string[];
  vibe: string[];
  intent: string;
  subtext: string;
  persona_trait: string[];
  bucket: string[];
}

## Technical Implementation Details

- **Language & Environment:** TypeScript on Node.js 18+
- **Embeddings:** Mocked 384-dim MiniLM vectors for semantic similarity
- **AI Replies:** GPT-like responses generated via mock API calls for cost-free development
- **Storage:** In-memory Maps simulating persistent storage keyed by `userId`
- **Testing:** Unit tests validate carry-in logic, profile updates, and pipeline integrity
- **Linting:** ESLint ensures code quality and maintainability

---

## How to Run

 Install dependencies:

   ```bash
   npm install
   npm run simulate:first
   npm run simulate:hundred