UVOM — Upgraded Version of Me

A Curated Multi-Agent AI Growth Companion

UVOM (Upgraded Version of Me) is an agentic AI learning orchestration platform engineered to solve the problem of passive content consumption. It replaces algorithmic dopamine loops with a continuous, personalized, measurable learning and personal-growth system.

While conventional social and recommendation platforms optimize primarily for attention and engagement, UVOM optimizes for human potential. It does this by continuously mapping a learner's aspirations, habits, evolving identity, skills, and progress to curate learning experiences that drive purposeful growth.

---

🎯 1. The Core Problem & Transformation Vision

The Problem Statement

"Learners consume more content but may learn less effectively."

Modern users are trapped in a cycle of passive consumption. Content recommendation algorithms optimize for watch-time, failing to address whether the media actually facilitates personal development. What’s lacking is continuous personalization, active assessment, and deep contextual understanding.

The UVOM Transformation Loop

UVOM establishes a stateful intelligence loop to flip this paradigm:

Traditional Platform Loop: User → Content → Watch → Recommendation → More Content (Attention Optimized)

UVOM Agentic Loop: User → Understand Goals → Build Skill Model → Identify Gaps → Curate Content → Generate Tasks → Assess Learning → Track Progress → Update Learner Model → Better Personalization (Growth Optimized)

---

🧠 2. Modular Multi-Agent Architecture (Google ADK)

The most significant technical innovation of UVOM is its reliance on a multi-agent orchestration architecture rather than a single, monolithic LLM prompt. Powered by the Google Agent Development Kit (ADK) + LlmAgent + Sequential Workflows, the system decomposes complex reasoning into specialized agents.

🤖 Root Orchestrator Agent

The central nervous system of UVOM. It coordinates workflows, maintains context, and determines which specialized agent capability should execute based on the user's immediate context.

👤 Profiler Agent (Learner Intelligence Layer)

Responsible for maintaining the continuously evolving Learner Model.

• Capabilities: Onboarding, Skill Mapping, Mood/State Tracking, Progress Evaluation, and Inactivity Detection.
• Mechanism: Computes the "Current Self" vs. "Future Self" dynamic to establish a concrete Skill-Gap Model (SkillGap = DesiredSkillProfile - CurrentSkillProfile).

🔭 Scout Agent (Discovery Engine)

Responsible for discovering highly relevant learning resources across the web that directly target the identified Skill Gaps.

• Capabilities: YouTube search, book search, event discovery, Instagram content curation, and article scraping.

📋 Task Master Agent (Execution Planner)

Converts high-level learning objectives into actionable, bite-sized tasks.

• Capabilities: Generates structured daily tasks, tracks task completion, and adapts learning paths dynamically when the learner becomes inactive or struggles.
• Example Output: "1. Watch curated communication video → 2. Practice 2-minute speech → 3. Complete quick quiz."

🔬 Content Analyzer Agent

UVOM does not just recommend links; it transforms raw media into structured learning data. It features distinct sub-analyzers:

• YouTube Analyzer: Uses APIs and extraction tools to process video transcripts.
• Instagram Analyzer: Interprets short-form educational content.
• Text/Book Analyzer: Processes written knowledge.

---

⚙️ 3. LLM Strategy: Model Routing

UVOM implements a model-routing strategy to optimize latency, inference cost, and reasoning depth:

• Gemini 2.5 Pro: Tasked with complex reasoning, deep skill-gap detection, Socratic tutoring, and architectural planning. Used when cognitive depth is required over speed.
• Gemini Flash: Handles fast intent parsing, real-time UI responses, and rapid quiz generation. Used for low-latency, transactional LLM interactions.

---

🛡️ 4. Tool Execution & Context via FastMCP

To ensure data integrity and prevent uncontrolled LLM actions (hallucinated SQL queries), UVOM utilizes the Model Context Protocol (MCP) and FastMCP.

• Controlled Capability Boundary: Agents do not interact with databases directly. Instead, they invoke strict, schema-validated tools.
• Shared Tool Layer: Includes capabilities such as:
  • youtube_data_api & yt_dlp_extract
  • google_search
  • whisper_transcribe & gemini_vision
  • embedding_tool & db_read/write
  • skill_matcher & progress_tracker

Example: Instead of arbitrary database writes, an agent executes a structured JSON tool call like {"tool": "update_learning_plan", "user_id": 42, "skill": "communication"}.

---

💾 5. Dual-Database Architecture

The backend explicitly separates transactional/structured application state from semantic knowledge, a best practice for modern RAG and agentic systems.

1. SQLite (Structured MVP Data)

Used for structured relational querying via SQLModel / SQLAlchemy ORM.

• Stores: User Profiles, User Skills, Daily Tasks, Activity Logs, and Telemetry.

2. ChromaDB (Semantic Vector Store)

Used for high-dimensional embedding storage and semantic similarity search.

• Stores: Skill embeddings, document chunks, and video transcript embeddings.
• Function: When a learner has a specific skill gap, the gap is vectorized to retrieve the most semantically relevant curated content.

---

⚡ 6. Event-Driven Learning Workflows

UVOM is not just a request-response chatbot; it is deeply event-driven.

Example Scenario: Video Checkpoint Assessment

1. The learner is watching a curated YouTube video on the React Frontend.
2. They reach a 10-minute checkpoint.
3. The frontend emits an event webhook to the FastAPI Backend Gateway.
4. FastAPI triggers a Google ADK Sequential Workflow.
5. The workflow executes an LLM pipeline to: Generate a contextual quiz based on the video transcript → Wait for the user's answer → Grade the answer → Update the Profiler Agent with new skill telemetry.

---

🏗️ 7. Full Technology Stack Summary

Architecture Layer	Technologies & Frameworks
Frontend / UI	React.js, TypeScript, Vite, Tailwind CSS, Lucide Icons
Backend API Gateway	Python, FastAPI
Database (Relational)	SQLite (MVP), PostgreSQL (Scale), SQLModel, SQLAlchemy
Database (Vector)	ChromaDB
Agent Orchestration	Google ADK, LlmAgent, Sequential Workflows
LLM Reasoning Layer	Gemini 2.5 Pro (Deep Reasoning), Gemini Flash (Low Latency)
Context & Tool Integration	FastMCP, Model Context Protocol (MCP)
Deployment / DevOps	Docker, Cloud-Native deployment ready

---

🚀 8. Future Scope & Scalability Vision

The architecture is built to be highly modular. New specialized agents can be integrated into the Root Orchestrator without modifying the core platform.

• Persistent Learner Memory: Maintaining long-term knowledge, strengths, weaknesses, and progress across multiple courses for lifelong personalized learning.
• Hyper-Personalization Engine: Continuously adapting learning paths, pace, memory, and performance dynamically.
• Interactive PDF & E-Book Learning: Allow learners to upload, annotate, summarize, and query study materials interactively.
• Voice-based AI Tutor & Viva Practice: Oral practice, natural conversations, and instant feedback via voice AI agents.
• Collaborative Peer Learning: AI-facilitated study groups, discussion rooms, and shared challenges.
• Multi-language Accessibility: Global inclusivity through localized learning experiences.

---

📌 UVOM in One Line:

React → FastAPI → Google ADK → Specialized Agents → MCP Tools → SQLite + ChromaDB → Personalized Learning Loop