Learn an AI Agent with LangChain & Gemini | Python + LangGraph | Autonomous Customer Support

https://www.youtube.com/watch?v=AgconCK-l4g

⇒ 𝐌𝐨𝐬𝐭 𝐨𝐟 𝐮𝐬 𝐭𝐡𝐢𝐧𝐤 𝐀𝐈 𝐜𝐮𝐬𝐭𝐨𝐦𝐞𝐫 𝐬𝐮𝐩𝐩𝐨𝐫𝐭 𝐢𝐬 𝐣𝐮𝐬𝐭 𝐚𝐧 𝐋𝐋𝐌 + 𝐚 𝐬𝐲𝐬𝐭𝐞𝐦 𝐩𝐫𝐨𝐦𝐩𝐭.

Actually, that approach may work for a demo, but production support needs much more.

When a customer asks to check an order, change a reservation, or request a refund, the system needs to manage 𝙨𝙩𝙖𝙩𝙚, 𝙩𝙤𝙤𝙡𝙨, 𝙥𝙚𝙧𝙢𝙞𝙨𝙨𝙞𝙤𝙣𝙨, 𝙫𝙖𝙡𝙞𝙙𝙖𝙩𝙞𝙤𝙣, 𝙖𝙣𝙙 𝙚𝙭𝙚𝙘𝙪𝙩𝙞𝙤𝙣.

A solid architecture looks like this:

✅ 𝙆𝙚𝙚𝙥 𝙨𝙩𝙖𝙩𝙚 𝙤𝙪𝙩𝙨𝙞𝙙𝙚 𝙩𝙝𝙚 𝙇𝙇𝙈: your application should manage session data, transactions, authentication, and tool results.

✅ 𝙐𝙨𝙚 𝙩𝙝𝙚 𝙇𝙇𝙈 𝙖𝙨 𝙖 𝙧𝙤𝙪𝙩𝙚𝙧: let the model understand intent, choose the right tool, and extract parameters.

For example:

𝚐𝚎𝚝_𝚘𝚛𝚍𝚎𝚛_𝚜𝚝𝚊𝚝𝚞𝚜(𝚘𝚛𝚍𝚎𝚛_𝚒𝚍)
𝚒𝚗𝚒𝚝𝚒𝚊𝚝𝚎_𝚛𝚎𝚏𝚞𝚗𝚍(𝚘𝚛𝚍𝚎𝚛_𝚒𝚍)

The backend should handle the actual database operations and business rules.

✅ 𝙆𝙚𝙚𝙥 𝙚𝙭𝙚𝙘𝙪𝙩𝙞𝙤𝙣 𝙙𝙚𝙩𝙚𝙧𝙢𝙞𝙣𝙞𝙨𝙩𝙞𝙘: tools should return structured results. Your runtime decides what happens next.

If validation fails, permissions are missing, or human intervention is required, your application should handle it with explicit rules.

The key separation is simple:

𝑳𝑳𝑴 → 𝒓𝒆𝒂𝒔𝒐𝒏𝒊𝒏𝒈 & 𝒓𝒐𝒖𝒕𝒊𝒏𝒈
𝑩𝒂𝒄𝒌𝒆𝒏𝒅 → 𝒃𝒖𝒔𝒊𝒏𝒆𝒔𝒔 𝒍𝒐𝒈𝒊𝒄 & 𝒆𝒙𝒆𝒄𝒖𝒕𝒊𝒐𝒏
𝑹𝒖𝒏𝒕𝒊𝒎𝒆 → 𝒔𝒕𝒂𝒕𝒆 & 𝒐𝒓𝒄𝒉𝒆𝒔𝒕𝒓𝒂𝒕𝒊𝒐𝒏

That separation is what makes an AI agent more predictable, auditable, and reliable in production.

An AI support agent isn't just a chatbot with a better prompt; 𝙄𝙩'𝙨 𝙖 𝙨𝙤𝙛𝙩𝙬𝙖𝙧𝙚 𝙨𝙮𝙨𝙩𝙚𝙢 𝙬𝙞𝙩𝙝 𝙖𝙣 𝙇𝙇𝙈 𝙞𝙣𝙨𝙞𝙙𝙚 𝙞𝙩.


▶️ I walk through how to build this kind of agent from scratch here:

https://www.youtube.com/watch?v=AgconCK-l4g

#AI #AIAgents #GenerativeAI #LLM #MachineLearning #Python #SoftwareEngineering #CustomerSupport #AIEngineering #Automation

𝐀𝐈 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 𝐜𝐨𝐦𝐞𝐬 𝐰𝐢𝐭𝐡 𝐚𝐧 𝐮𝐧𝐜𝐨𝐦𝐟𝐨𝐫𝐭𝐚𝐛𝐥𝐞 𝐫𝐞𝐚𝐥𝐢𝐭𝐲: 𝐲𝐨𝐮𝐫 𝐜𝐨𝐝𝐞 𝐜𝐚𝐧 𝐛𝐫𝐞𝐚𝐤 𝐞𝐯𝐞𝐧 𝐰𝐡𝐞𝐧 𝐲𝐨𝐮𝐫 𝐥𝐨𝐠𝐢𝐜 𝐢𝐬 𝐜𝐨𝐫𝐫𝐞𝐜𝐭.

I have experienced this firsthand while building AI agents with Gemini and LangChain.

➜ A model endpoint changes.

➜ A parameter gets renamed.

➜ A framework updates its API.

A response that used to be a string becomes a structured object.

Suddenly, perfectly reasonable code starts throwing errors.

What I have learned from that:

✅ 𝑫𝒐𝒏’𝒕 𝒕𝒊𝒈𝒉𝒕𝒍𝒚 𝒄𝒐𝒖𝒑𝒍𝒆 𝒚𝒐𝒖𝒓 𝒂𝒑𝒑𝒍𝒊𝒄𝒂𝒕𝒊𝒐𝒏 𝒕𝒐 𝒇𝒓𝒂𝒎𝒆𝒘𝒐𝒓𝒌 𝒊𝒏𝒕𝒆𝒓𝒏𝒂𝒍𝒔: Keep your business logic separate from model and framework integrations.

✅ 𝑬𝒙𝒑𝒆𝒄𝒕 𝑨𝑷𝑰𝒔 𝒕𝒐 𝒆𝒗𝒐𝒍𝒗𝒆: Pin important dependencies, read changelogs, and test upgrades before pushing them into production.

✅ 𝑵𝒆𝒗𝒆𝒓 𝒂𝒔𝒔𝒖𝒎𝒆 𝒎𝒐𝒅𝒆𝒍 𝒐𝒖𝒕𝒑𝒖𝒕 𝒉𝒂𝒔 𝒐𝒏𝒆 𝒇𝒐𝒓𝒎𝒂𝒕: Structured responses, tool calls, metadata, and plain text can all require different parsing strategies.

✅ 𝑩𝒖𝒊𝒍𝒅 𝒔𝒎𝒂𝒍𝒍 𝒊𝒏𝒕𝒆𝒈𝒓𝒂𝒕𝒊𝒐𝒏 𝒍𝒂𝒚𝒆𝒓𝒔: If Gemini or LangChain changes, you should be able to update one part of your application instead of rewriting the entire agent.

The goal isn't to avoid change, but to make change cheap. While deploying my latest AI agent with Gemini and LangChain, I ran into several of these exact edge cases.

I documented the fixes and the complete setup here:

👉 https://www.youtube.com/watch?v=AgconCK-l4g

If you're building AI agents with Python, this is one lesson worth learning early.

#AIEngineering #AIAgents #Python #LangChain #Gemini #LLM #GenerativeAI #SoftwareEngineering #MachineLearning

When I build an AI agent, I do not start by asking, Which model should I use? I start by designing the system around the model.



The model provides reasoning and language capabilities. The surrounding architecture determines whether the agent is reliable, controllable, and production ready.



This is the approach I follow:

𝟏. 𝐌𝐨𝐝𝐞𝐥: I select the model based on reasoning capability, task complexity, latency, cost, and context requirements.

𝟐. 𝐓𝐨𝐨𝐥𝐬: I give the agent well-defined tools with strict schemas, validation, permissions, and predictable outputs.

𝟑. 𝐂𝐨𝐧𝐭𝐞𝐱𝐭: I carefully control the information provided to the model through retrieval, memory, conversation state, and structured context.

𝟒. 𝐎𝐫𝐜𝐡𝐞𝐬𝐭𝐫𝐚𝐭𝐢𝐨𝐧: I define how the agent reasons, when it can call tools, when it should retry, when it should ask for clarification, and when it must stop.

𝟓. 𝐆𝐮𝐚𝐫𝐝𝐫𝐚𝐢𝐥𝐬: I validate inputs, tool calls, and outputs. For sensitive or high-impact operations, I add additional verification.

𝟔. 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐛𝐢𝐥𝐢𝐭𝐲: I monitor tool calls, model responses, latency, failures, token usage, and agent execution paths.

𝟕. 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧: I test the complete system against realistic scenarios, edge cases, adversarial inputs, and expected failure modes.

▶️ I walk through how to build this kind of agent from scratch here:


https://www.youtube.com/watch?v=AgconCK-l4g



#AIEngineering #AIAgents #GenerativeAI #LLM #MachineLearning #Python #LangChain #LangGraph #MLOps #SoftwareEngineering

FastAPI From Zero: Build a Production AI API | Episode 1 - Course Overview
https://www.youtube.com/watch?v=0SLLG2Z_Htw

When I build an AI application, choosing the backend framework is an important decision.

There are several good options, but I usually look at 𝐅𝐚𝐬𝐭𝐀𝐏𝐈, 𝐃𝐣𝐚𝐧𝐠𝐨, and 𝐅𝐥𝐚𝐬𝐤 first.

The choice really depends on what I'm building.

✔️ 𝐅𝐚𝐬𝐭𝐀𝐏𝐈 makes a lot of sense when the application is mainly an AI/API backend. Since most AI tools I use are already in Python, I can keep the whole stack in one ecosystem, from LLMs and embeddings to document processing, RAG, databases, and the API itself.

✔️ 𝐃𝐣𝐚𝐧𝐠𝐨 is a strong choice when the AI functionality is part of a larger web application. Its built-in ORM, authentication, admin panel, and other features can save a lot of development time.

✔️ 𝐅𝐥𝐚𝐬𝐤 is still a great option when I want something simple, lightweight, and flexible, especially for smaller services or prototypes.

For an AI application, I also need to think beyond the framework:

✔️ Authentication

✔️ Database and data persistence

✔️ Document processing

✔️ Embeddings and vector search

✔️ RAG

✔️ Background tasks

✔️ Testing

✔️ Docker

✔️ Monitoring

✔️ Deployment

There isn't one framework that is "best" for every AI application. For the type of production AI backends I'm building, 𝐅𝐚𝐬𝐭𝐀𝐏𝐈 is often a practical choice because it provides a clean API layer while keeping everything close to the Python AI ecosystem.

The framework is only one piece of the puzzle.

Good architecture matters more than the framework you choose.


What do you normally use for AI applications: 𝐅𝐚𝐬𝐭𝐀𝐏𝐈, 𝐃𝐣𝐚𝐧𝐠𝐨, 𝐅𝐥𝐚𝐬𝐤, or something else?

Here is the roadmap to build an AI application with FastAPI: https://www.youtube.com/watch?v=0SLLG2Z_Htw


#FastAPI #Python #AIEngineering #GenerativeAI #RAG #BackendDevelopment #MachineLearning #Django #Flask #SoftwareArchitecture

FastAPI Fundamentals: Build Your First AI API | Python FastAPI Course (Episode 2 - Overview of API)
https://youtu.be/vvP9GIWSews

𝐅𝐚𝐬𝐭𝐀𝐏𝐈 𝐯𝐬 𝐑𝐄𝐒𝐓 𝐀𝐏𝐈 — What’s the Difference?

One thing I see quite often when people start building APIs with Python is confusion between FastAPI and REST API.

In reality, they are not the same thing.

𝐑𝐄𝐒𝐓 𝐀𝐏𝐈 is an architectural approach for designing APIs around resources, HTTP methods, stateless communication, and standard HTTP responses.

𝐅𝐚𝐬𝐭𝐀𝐏𝐈 is a Python web framework that helps you build APIs.

For example, in an AI application, I might have:

GET /documents
𝙶𝙴𝚃 /𝚍𝚘𝚌𝚞𝚖𝚎𝚗𝚝𝚜
𝙿𝙾𝚂𝚃 /𝚍𝚘𝚌𝚞𝚖𝚎𝚗𝚝𝚜
𝙶𝙴𝚃 /𝚍𝚘𝚌𝚞𝚖𝚎𝚗𝚝𝚜/{𝚒𝚍}
𝙿𝚄𝚃 /𝚍𝚘𝚌𝚞𝚖𝚎𝚗𝚝𝚜/{𝚒𝚍}
𝙳𝙴𝙻𝙴𝚃𝙴 /𝚍𝚘𝚌𝚞𝚖𝚎𝚗𝚝𝚜/{𝚒𝚍}

These endpoints can follow 𝐑𝐄𝐒𝐓 principles.

𝐅𝐚𝐬𝐭𝐀𝐏𝐈 is the tool I use to implement them in Python.

So, a simple way to remember it:

𝐑𝐄𝐒𝐓 = how the API is designed
𝐅𝐚𝐬𝐭𝐀𝐏𝐈 = the framework used to build it

𝐅𝐚𝐬𝐭𝐀𝐏𝐈 also gives us useful features such as request validation, automatic API documentation, dependency injection, and strong support for asynchronous applications.

Understanding this distinction makes it much easier to understand 𝐅𝐚𝐬𝐭𝐀𝐏𝐈 and, more importantly, to design APIs properly.

𝐅𝐚𝐬𝐭𝐀𝐏𝐈 Fundamentals: Build Your First AI API | Python FastAPI Course (An Overview of API): https://youtu.be/vvP9GIWSews

#FastAPI #Python #RESTAPI #APIDevelopment #AI #MachineLearning #BackendDevelopment

𝐁𝐞𝐟𝐨𝐫𝐞 𝐜𝐡𝐚𝐧𝐠𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐌𝐋 𝐦𝐨𝐝𝐞𝐥, 𝐜𝐡𝐞𝐜𝐤 𝐲𝐨𝐮𝐫 𝐝𝐚𝐭𝐚.
When a model performs badly, the first thing we often do is try a different algorithm.

Sometimes that works.

But before doing that, I usually look at the dataset. 🔍

I check things like:

🔹 Missing values
🔹 Duplicate records
🔹 Outliers
🔹 Wrong data types
🔹 Class imbalance
🔹 Data leakage
🔹 High-cardinality columns
🔹 Features with little useful information

There is no point spending hours tuning a model if the dataset itself has problems. ⚠️

A simple workflow I prefer is:

📥 𝑹𝒂𝒘 𝑫𝒂𝒕𝒂
↓
🔎 𝑪𝒉𝒆𝒄𝒌 𝑸𝒖𝒂𝒍𝒊𝒕𝒚
↓
🧹 𝑪𝒍𝒆𝒂𝒏
↓
📊 𝑨𝒏𝒂𝒍𝒚𝒛𝒆
↓
🤖 𝑻𝒓𝒂𝒊𝒏
↓
📈 𝑴𝒐𝒏𝒊𝒕𝒐𝒓

Data quality is not just something to deal with before machine learning. It affects every step that comes after it.

So when a model is not performing as expected, don't immediately change the model.

🔍 Take another look at the data first. https://youtube.com/playlist?list=PL0nX4ZoMtjYHTtowSzzB2gVH2AuuoF9WW&si=EhLKvJCVlYQXknOs
Use Data quality checker tool: https://datasetdoctor.fastapicloud.dev

What is FastAPI? https://www.youtube.com/watch?v=yUsDgLZPDyI

Your model can look excellent and still be wrong.

One of the first things I check when evaluating an ML dataset is data leakage. 🔍

Data leakage happens when information that would not actually be available at prediction time gets into the training data.

For example:

🏥 Healthcare
You are predicting whether a patient will be admitted, but your dataset includes a field recorded after admission.

💳 Fraud detection
You are predicting fraud, but one of the features is created after the transaction has already been investigated.

📦 Customer churn
You are predicting who will leave, but the training data contains information that only becomes available after the customer leaves.

The result?

Your model may show:

📈 98% accuracy
📈 Excellent validation results
📈 Great performance during testing

Then you put it into production...

And the performance drops.

The problem was not necessarily the model.

The model had access to information it would never have in the real world.

That is why I don't look at model performance alone.

I also ask:

🔎 Where did each feature come from?
⏱️ When was it created?
🎯 Would this information actually be available when making the prediction?

A high score is not always a good score.

Sometimes, it is a warning sign.

Check out data quality issues
https://youtube.com/playlist?list=PL0nX4ZoMtjYHTtowSzzB2gVH2AuuoF9WW&si=EhLKvJCVlYQXknOs

Also checkout data quality checker tool https://datasetdoctor.fastapicloud.dev

#MachineLearning #DataScience #AI #DataLeakage #MLOps #Python

Set Up FastAPI Development Environment with uv & VS Code | FastAPI Full Course(Episode 4)
https://www.youtube.com/watch?v=G60LwkySnwQ

FastAPI Full Course Episode 5: FastAPI Parameters & Request Bodies(Path, Query & Pydantic) https://www.youtube.com/watch?v=-tkww4I4Vfg&t=638s

FastAPI Episode 6: Pydantic Validation Requests and Response Models
https://youtu.be/G5cKA88-6vc

The hardest part of building AI applications isn't writing the prompt or calling the model. In the last two weeks, I learned that keeping the backend from turning into spaghetti code once you move past the tutorial phase.

When you're wiring up an AI document pipeline in FastAPI, a few things quickly become non-negotiable:



• Payload Guardrails: If your Pydantic schemas aren't catching malformed JSON, missing nested fields, or bad Enums at the door, your AI service will fail unpredictably downstream.



• Route Isolation: Mixing your raw API endpoints with validation logic and business rules makes refactoring a nightmare by week three.



• The Persistence Gap: Transitioning from mock in-memory data structures to a real relational database and a vector store for RAG is where most clean prototypes start to break down.



If you're building production backends for AI and ML features, where do you usually draw the line between keeping things simple and over-engineering your architecture?



You can learn about FastAPI: https://www.youtube.com/playlist?list=PLQNCas8_eikM



#FastAPI #Python #BackendEngineering #SoftwareArchitecture #APIs #Pydantic #ArtificialIntelligence #MachineLearning #RAG

FastAPI Episode 7: Advanced Pydentic Model Design
https://www.youtube.com/watch?v=j4lLM6tWKKk

When building a FastAPI application, I pay close attention to how data is validated before it reaches the business logic.

That is where Pydantic becomes particularly useful.

Pydantic lets us define the structure and rules for the data our application accepts, instead of scattering validation checks throughout the codebase.

I use it to define and enforce the data contract at the application boundary.

For example:

𝑐𝑙𝑎𝑠𝑠 𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔𝐶𝑜𝑛𝑓𝑖𝑔(𝐵𝑎𝑠𝑒𝑀𝑜𝑑𝑒𝑙):

𝑐ℎ𝑢𝑛𝑘_𝑠𝑖𝑧𝑒: 𝑖𝑛𝑡 = 𝐹𝑖𝑒𝑙𝑑(𝑔𝑡=0)

𝑜𝑣𝑒𝑟𝑙𝑎𝑝: 𝑖𝑛𝑡 = 𝐹𝑖𝑒𝑙𝑑(𝑔𝑒=0)

@𝑚𝑜𝑑𝑒𝑙_𝑣𝑎𝑙𝑖𝑑𝑎𝑡𝑜𝑟(𝑚𝑜𝑑𝑒="𝑎𝑓𝑡𝑒𝑟")

𝑑𝑒𝑓 𝑣𝑎𝑙𝑖𝑑𝑎𝑡𝑒_𝑐𝑜𝑛𝑓𝑖𝑔(𝑠𝑒𝑙𝑓):

𝑖𝑓 𝑠𝑒𝑙𝑓.𝑜𝑣𝑒𝑟𝑙𝑎𝑝 >= 𝑠𝑒𝑙𝑓.𝑐ℎ𝑢𝑛𝑘_𝑠𝑖𝑧𝑒:

𝑟𝑎𝑖𝑠𝑒 𝑉𝑎𝑙𝑢𝑒𝐸𝑟𝑟𝑜𝑟("𝑜𝑣𝑒𝑟𝑙𝑎𝑝 𝑚𝑢𝑠𝑡 𝑏𝑒 𝑠𝑚𝑎𝑙𝑙𝑒𝑟 𝑡ℎ𝑎𝑛 𝑐ℎ𝑢𝑛𝑘_𝑠𝑖𝑧𝑒")

𝑟𝑒𝑡𝑢𝑟𝑛 𝑠𝑒𝑙𝑓

Both fields can be individually valid, while their combination is not.

That is the difference between:

Field validation → Is this value valid?

Model validation → Is this combination valid?

With Pydantic, 𝙁𝙞𝙚𝙡𝙙(), 𝙛𝙞𝙚𝙡𝙙_𝙫𝙖𝙡𝙞𝙙𝙖𝙩𝙤𝙧(), 𝙖𝙣𝙙 𝙢𝙤𝙙𝙚𝙡_𝙫𝙖𝙡𝙞𝙙𝙖𝙩𝙤𝙧() let us keep these rules close to the schema.

The result is a cleaner boundary:

𝙍𝙚𝙦𝙪𝙚𝙨𝙩 → 𝙑𝙖𝙡𝙞𝙙𝙖𝙩𝙞𝙤𝙣 → 𝘽𝙪𝙨𝙞𝙣𝙚𝙨𝙨 𝙇𝙤𝙜𝙞𝙘 → 𝘿𝙖𝙩𝙖𝙗𝙖𝙨𝙚 / 𝘼𝙄

For production AI applications, this matters. Document metadata, processing parameters, search filters, and structured AI outputs all need predictable contracts.

Good schemas do more than describe data. They protect the rest of the system.


You can explore more: https://www.youtube.com/playlist?list=PLQNCas8_eikM


#Python #Pydantic #PydanticV2 #FastAPI #BackendEngineering #AIEngineering

✅ FastAPI Episode 8: Organizing Routes with APIRouter
https://youtu.be/SGPXI7bzEDM

This is the github starter for the project based FastAPI course tutorial
https://github.com/epythonlab2/ai-document-api