AI Learning Roadmap — Part 3

Advanced Level

1. Model Context Protocol (MCP)

A standard for AI to connect to your tools and data. Instead of writing custom integrations for each system, MCP acts as a universal interface.

Problem it solves:

**Problem
**User asks:“Summarize the client’s complaint from email, check the related Jira ticket, and draft a reply in Slack.”

Your agent must:
Know Gmail API
Know Jira API
Know Slack API
Know how to connect the data between them
Know how to authenticate all of them
Know how to format everything

**Solution:
**With MCP, all these tools expose themselves in a standard way.

Your agent doesn’t care if it’s Gmail, Jira, or Notion.

It just says:
“Get the latest email from client”
“Find related ticket”
“Post message to Slack”

Example workflow:

Step 1: You create an MCP server for your tool

`MCP Server for MyDatabase:

Define what the AI can ask for

Define how to handle those requests

Handle the database call

Return results to AI`

Step 2: Claude connects to your MCP server

Claude: "Get the user with email john@example.com"

MCP Server: Queries database

MCP Server: Returns user data

Claude: Uses that data to answer the user

Why it matters: MCP solves the problem of AI agents needing custom integrations for every tool by giving them one universal way to talk to all tools.

2. Agents and Tool Use

AI agents can decide which tools to use and act on your behalf, not just answer questions.

Simple analogy:

Without agents: AI answers “What’s the weather?”

With agents: AI checks weather, calendar, to-do list, then plans your day.

Example:

Without agents:

User: “Book a meeting with Sarah for next Tuesday”

AI: “I can’t book meetings. You need to use your calendar app.”

Result: User has to do the work.

With agents:

User: “Book a meeting with Sarah for next Tuesday”

AI: 1. Uses calendar tool to check your availability Tuesday

2. Uses contact tool to find Sarah’s email

3. Uses email tool to send meeting invitation

4. Confirms: “Done! Sent meeting invite to Sarah for Tuesday 2pm”

Result: Task completed automatically.

How agents work:

`Step 1: Understand the goal User: “What’s the weather and how busy am I today?”

Step 2: AI decides what tools to use “I need: Weather tool + Calendar tool”

Step 3: AI uses each tool “Get weather for [location]” “Get calendar events for [today]”

Step 4: Get results back Weather tool: “Sunny, 75°F” Calendar tool: “2 meetings scheduled, 4 hours free”

Step 5: Synthesize answer AI: “It’s sunny and 75°F today. You have 2 meetings and 4 hours of free time.”`

Why it matters: Agents automate complex workflows, saving huge amounts of manual work.

3. Vector Databases and Similarity Search

A specialized database that stores embeddings (numeric representations of text) and finds similar ones quickly.

Why it’s needed: Keyword search misses documents with different wording but same meaning.

Example:

Regular database (doesn’t work well):

Search for claims with matching keywords “car damaged by collisions” might not find similar claim with different wording like “Vehicle struck by impact”

Vector database (works great):

Convert new claim to embedding.
Search vector database for similar embeddings.
Instantly finds all related claims, no matter the wording.

Why it matters: Vector DBs are essential for semantic search, RAG systems, and scalable AI applications.

4. Streaming and Chunking

Streaming: Getting AI responses word-by-word in real-time instead of waiting for the full response
Chunking: Breaking big documents into smaller pieces before processing

Streaming example:

Without streaming:

User: “Write a 1000 word essay…”
[waiting 8 seconds…]
Result: Entire essay appears at once. Feels slow to users.

With streaming:

User: “Write a 1000 word essay”
Result: Words start appearing immediately. Users see progress in real time

Why streaming matters:

Users feel like AI is thinking in real-time
Better user experience
Modern apps expect streaming

Chunking example:

You have a 100-page technical manual (200,000 tokens). The AI’s context window is 100,000 tokens.

Problem:

Try to load whole manual Exceeds context window AI can't process it

Solution (chunking):

`Split manual into 5 chunks of 40,000 tokens each

User asks: “How do I reset the device?” System finds which chunk has reset instructions Loads just that chunk (40,000 tokens) AI answers based on that chunk Works perfectly!`

Chunking strategies:

Simple: fixed token count → may split mid-sentence
Smart: split at sentence/paragraph → preserves context
Overlap: chunks overlap → prevents info loss at boundaries

Why it matters: Improves speed, usability, and accuracy in production AI systems.

5. Cost Optimization (Input vs Output Tokens)

Key Terms: Input Tokens, Output Tokens, Token Economy

AI APIs charge per token. Understanding token usage saves money.

Input tokens (your prompt + context)
Output tokens (AI’s response)

Input is usually cheaper than output.

Example pricing:

Input: $0.003 per 1,000 tokens ($3 per million)

Output: $0.009 per 1,000 tokens ($9 per million)

Discount for cached input: 90% cheaper

You send: 1,000 input tokens + get 200 output tokens back

Cost: $0.003 + $0.0018 = $0.0048 (about half a cent)

Optimization strategies:

Reduce input size: Include only relevant parts of documents
Caching: Store repeated context → 90% discount on reused tokens
Cheaper models for simple tasks: Use smaller LLMs for easy jobs
Batch processing: Process multiple items together → fewer API calls

Why it matters: Token add up quickly at scale. Optimizing saves up on costs.