🚀 90 Days of AI Engineering

From "I can code a little" to shipping production AI products.

A free, rigorous, day-by-day roadmap engineered to build real-world AI applications.
No expensive bootcamps. No endless passive video watching. Every week ends with a deployed project.

Star This Repo ⭐ • Fork & Track Progress 🍴 • Join Discussions 💬 • Weekly Projects 🛠️

🌟 Why This Roadmap?

Most AI tutorials fall into two traps:

Too theoretical: 40 hours of linear algebra before you get to invoke an API.
Too shallow: Copy-pasting a 5-line Streamlit wrapper around an OpenAI API key without understanding tokens, chunking, evals, or rate limits.

AI Engineering is different. It is about building reliable, maintainable software systems with non-deterministic models:

🔗 Grounding models in custom data (RAG, Vector Search, Hybrid Retrieval, Rerankers)
🤖 Empowering LLMs with deterministic capabilities (Function Calling, Tool Use, MCP, Agents)
📏 Measuring quality and preventing regressions (Automated Evals, LLM-as-a-judge, CI/CD)
⚡ Deploying performant, production backends (FastAPI streaming, Docker, Token counting, Cost caps)

Every single week ends with code you wrote and shipped, not just videos you watched.

🗺️ Roadmap Architecture

flowchart TD
    subgraph Phase1 [Phase 1: Foundations & Mechanics]
        D1["Days 1-14: Python, Git & APIs"] --> D15["Days 15-21: How LLMs Work"]
        D15 --> D22["Days 22-28: Prompting & Structured Outputs"]
    end

    subgraph Phase2 [Phase 2: Full-Stack AI Services]
        D22 --> D29["Days 29-35: Backend with FastAPI"]
        D29 --> D36["Days 36-42: Embeddings & Vector Search"]
        D36 --> D43["Days 43-56: Production RAG (2 Weeks)"]
    end

    subgraph Phase3 [Phase 3: Production, Evals & Agents]
        D43 --> D57["Days 57-63: LLM Evals & CI/CD"]
        D57 --> D64["Days 64-70: Tool Use & Autonomous Agents"]
        D64 --> D71["Days 71-77: MCP & AI Coding Workflows"]
        D71 --> D78["Days 78-84: Deploy, Secure & Monitor"]
    end

    subgraph Phase4 [Phase 4: Capstone & Career]
        D78 --> D85["Days 85-90: Capstone AI Product Launch"]
        D85 --> D91["Days 91-95: Portfolio, Resume & Career"]
    end

    style Phase1 fill:#161b22,stroke:#58a6ff,stroke-width:2px,color:#fff
    style Phase2 fill:#161b22,stroke:#bc8cff,stroke-width:2px,color:#fff
    style Phase3 fill:#161b22,stroke:#39d353,stroke-width:2px,color:#fff
    style Phase4 fill:#161b22,stroke:#f0883e,stroke-width:2px,color:#fff

📚 Table of Contents

📘 Introduction
👥 Who This Is For
🚀 Start Here
💰 Zero-Cost Guarantee
🛠️ Weekly Projects Overview
🗓️ Daily Curriculum
🔬 Optional Track: Under the Hood (Math & Deep Learning)
👨‍💻 Creator & Maintainer
🌍 Translations
🤝 Contributing
📜 License

📘 Introduction

AI engineering is about building products with models: calling LLMs, grounding them in your data, giving them tools, measuring quality, and running them in production. You do not need a PhD or a research background to do it.

This repository gives you a structured 90-day path. Each module includes:

📖 Learn: A curated list of free, well-regarded resources.
🛠️ Build: One project that proves you understood the concepts.
✅ Done When: A concrete checklist so you know with confidence when to move forward.

👥 Who This Is For

Students and self-taught developers who want to land an AI engineering job.
Software engineers transitioning from traditional backend, frontend, or full-stack roles into AI.
Builders & hackers who can write basic code and want a structured, no-fluff roadmap to master the modern AI stack.

Note

You do not need prior AI or data science experience. Days 1–14 cover programming, Git, and API fundamentals.

🚀 Start Here

Time budget: 1–2 focused hours a day. Day numbers are a guide, not a strict deadline. Consistency beats intensity.
Follow the order: Each week intentionally builds on top of the code you wrote in the previous project.
Build in public: Create your own GitHub repository for your daily work. Commit daily. Share your progress using #90DaysOfAIEngineering.
Track progress: Click Fork at the top right of this repo, copy our PROGRESS_TRACKER.md, and tick off each day as you finish it.
Community: Got questions or looking for study buddies? Jump into GitHub Discussions. Use Issues only for broken links and content improvements.

💰 Zero-Cost Guarantee

You do not need to pay for courses or spend hundreds on API credits:

Free Provider Credits: Most major LLM providers offer generous free tiers or trial credits (e.g., Google AI Studio Gemini API offers free rate limits, Groq offers high-speed free tier access, GitHub Models offers free API endpoints).
Run Local LLMs: Use Ollama or vLLM to run models like Llama 3, Mistral, or Qwen locally at $0 cost on your laptop or desktop.
Free Vector Databases: Chroma (embedded local), SQLite with sqlite-vec, or free tier cloud vector DBs (Pinecone, Qdrant).
Open-Source Tools: All courses, documentation, frameworks, and deployment platforms referenced in this guide have completely free tiers.

Caution

Never commit API keys to version control. Use .env files and add .env to your .gitignore.

🛠️ Weekly Projects Overview

Every week you will ship a production-grade component:

Phase	Days	Project Title	Key Tech	Starter Kit
01	Day 1–14	CLI LLM File Summarizer	Python, `httpx`, Git, APIs	`projects/day-01-14-cli-summarizer`
02	Day 15–21	Chat CLI with Token & Cost Tracker	`tiktoken`, Local Ollama / API	`projects/day-15-21-token-cost-chat`
03	Day 22–28	Validated Structured Data Extractor	Pydantic, Instructor / JSON Schema	`projects/day-22-28-structured-extractor`
04	Day 29–35	Production FastAPI Service	FastAPI, SSE Streaming, `pytest`, CI	`projects/day-29-35-fastapi-backend`
05	Day 36–42	Semantic Vector Search Engine	ChromaDB / pgvector, Embeddings	`projects/day-36-42-semantic-vector-search`
06	Day 43–56	Production RAG with Source Citations	Chunking, Hybrid Search, Reranking	`projects/day-43-56-production-rag`
07	Day 57–63	Automated LLM Evaluation Harness	Golden Datasets, LLM-as-a-Judge, CI	`projects/day-57-63-evals-ci`
08	Day 64–70	Autonomous Tool-Using Agent	Function Calling, Guardrails, Loops	`projects/day-64-70-autonomous-agent`
09	Day 71–77	Custom Model Context Protocol (MCP) Server	MCP SDK, JSON-RPC, AI Assistants	`projects/day-71-77-mcp-server`
10	Day 78–84	Secure Dockerized AI Deployment	Docker, OWASP Top 10, Cloud / Bedrock	`projects/day-78-84-deploy-docker-bedrock`
11	Day 85–90	Full-Stack AI Capstone Product	End-to-End System, Live Demo, CI	`projects/day-85-90-capstone-product`

🗓️ Daily Curriculum

Day 1–14: Python, Git and APIs

🎯 Focus

Building rock-solid software development fundamentals: modern Python 3, asynchronous programming, version control with Git, and interacting with REST APIs.

📖 Learn

Python Fundamentals: The Python Tutorial (Official) or CS50's Introduction to Programming with Python
Git & GitHub: Learn Git Branching (Interactive)
Environment & CLI: Terminal basics, virtual environments (venv), package managers (pip), and .env files with python-dotenv.
HTTP & JSON: REST conventions, HTTP methods, headers, status codes, and HTTP clients (requests or httpx).
Async Python: async / await, event loops, and why async I/O is critical for LLM applications.

🛠️ Build

Project: A command-line tool that reads a local text file, sends it to an LLM API (OpenAI, Anthropic, Gemini, or local Ollama), and streams/prints an executive summary.

📁 Starter Code & Specification

✅ Done When

Your project lives in a public Git repository with a clear README.md.
API keys are strictly loaded from environment variables—never committed.
You can explain what a request, response, payload, and HTTP status code are.
Your script handles file-not-found and API errors gracefully without crashing.

Day 15–21: How LLMs Work

🎯 Focus

Demystifying the underlying mechanics of Large Language Models: transformer architecture intuition, tokenization, context windows, sampling parameters, and cost accounting.

📖 Learn

Neural Networks Intuition: Andrej Karpathy: Neural Networks: Zero to Hero (watch the intro talks first, coding lessons later).
Transformers Visualized: 3Blue1Brown: Neural Networks Series (including the transformer and attention chapters).
LLM Primer: Hugging Face LLM Course (first chapters).
Core Concepts:
- Tokenization (Byte-Pair Encoding, token-to-word ratios).
- Context window limits and attention degradation.
- Sampling hyperparameters: Temperature, Top-P, Frequency/Presence penalties.
- Hallucination causes and cost estimation per 1k / 1M tokens.

🛠️ Build

Project: An interactive Terminal Chat CLI featuring persistent conversation memory and a live token count and dollar cost counter for every prompt/response turn.

📁 Starter Code & Specification

✅ Done When

You can explain tokens, BPE, and context windows to a peer without jargon.
You understand why identical prompts produce different outputs when temperature > 0.
Your CLI accurately tracks prompt tokens, completion tokens, and dollar expenditures across multiple providers.

Day 22–28: Prompting and Structured Outputs

🎯 Focus

Transitioning from unstructured text generation to deterministic, machine-readable structured JSON outputs using schema validation, system prompts, and few-shot examples.

📖 Learn

Prompt Engineering Guides: Anthropic Interactive Prompt Engineering Course
Official Documentation:
Validation Libraries: Pydantic v2 for data schemas and Instructor for structured extraction.

🛠️ Build

Project: A resilient data extractor that takes messy unstructured inputs (invoices, emails, receipts, or job postings) and produces validated, type-safe JSON with automatic retries on validation failure.

📁 Starter Code & Specification

✅ Done When

Tested against 20+ diverse, messy input samples with an automated pass-rate metric recorded.
Pydantic schema validation failures trigger structured retry loops with corrective feedback to the model.
The pipeline never crashes on malformed inputs; failures are logged cleanly.

Day 29–35: Backend with FastAPI

🎯 Focus

Packaging AI logic into production-grade, asynchronous web APIs with authentication, background processing, streaming responses, and automated test suites.

📖 Learn

FastAPI Fundamentals: FastAPI Official Tutorial
Advanced API Patterns: Server-Sent Events (SSE) for token streaming, dependency injection, middleware, and background tasks.
Testing & Quality: Writing unit and integration tests with pytest and httpx.AsyncClient.
CI/CD: Automating tests with GitHub Actions.

🛠️ Build

Project: A production-ready FastAPI service exposing:

A streaming chat endpoint using Server-Sent Events (SSE).

An async structured extraction endpoint.

API-key header authentication and rate limiting.

Comprehensive pytest test suite running in GitHub Actions.

📁 Starter Code & Specification

✅ Done When

All pytest suites run and pass automatically on push/PR via GitHub Actions CI.
The service can be spun up with a single command (uvicorn main:app --reload).
Token streaming works smoothly in both curl and frontend clients without buffering delays.

Day 36–42: Embeddings and Vector Search

🎯 Focus

The math and mechanics of representing text as dense vector embeddings, calculating vector similarities, and querying vector databases.

📖 Learn

Embeddings & Semantic Search: Hugging Face LLM Course - Embeddings Section
Vector Search Concepts: Pinecone Learning Center: Vector Search
Vector Databases: Hands-on with ChromaDB, Qdrant, or PostgreSQL with pgvector.
Similarity Metrics: Cosine similarity, dot product, and Euclidean distance.

🛠️ Build

Project: A semantic search CLI and Python library over a local folder of PDF, Markdown, and text documents using a local vector store (ChromaDB or sqlite-vec).

📁 Starter Code & Specification

✅ Done When

You can document and demonstrate a specific query where semantic search outperforms keyword search (BM25).
You can demonstrate a query where keyword search wins (e.g., exact part numbers or acronyms).
Chunking logic handles document boundaries cleanly without cutting sentences in half.

Day 43–56: RAG (Two Weeks)

🎯 Focus

Deep dive into Retrieval-Augmented Generation (RAG)—the core requirement in modern AI engineering jobs. Moving beyond naive RAG to advanced chunking, hybrid retrieval, reranking, and citation generation.

📖 Learn

RAG Architecture: RAG From Scratch Series (LangChain)
DeepLearning.AI Courses: Building Systems with the ChatGPT API & Advanced RAG Courses
Advanced Retrieval Techniques:
- Recursive character vs. semantic chunking.
- Hybrid search (combining BM25 keyword matching with dense vector retrieval).
- Cross-encoder reranking (e.g., Cohere Rerank or FlashRank).
- Mitigating hallucinations with strict context prompting.

🛠️ Build

Project: A production RAG system over your own knowledge base that returns high-accuracy answers with inline source citations ([Doc 1, Page 4]), and explicitly replies "I do not have enough information to answer this based on the provided documents" when information is missing.

📁 Starter Code & Specification

✅ Done When

You evaluated at least two distinct chunking strategies (e.g. fixed 500-token vs. recursive semantic) and documented accuracy trade-offs.
Every generated claim links directly to its source context passage.
The system achieves 0% hallucination on out-of-domain / unanswerable questions.

Day 57–63: Evals

🎯 Focus

The difference between a weekend toy and a production AI system is evals. Learn how to systematically measure model quality, benchmark prompts, detect regressions, and build CI evaluation gates.

📖 Learn

Eval Principles: Hamel Husain's Guide to Evals
Evaluation Methodologies: Unit testing for LLMs, golden evaluation datasets, and LLM-as-a-judge (strengths and blindspots).
Frameworks: Ragas, DeepEval, or writing lightweight custom eval runners.

🛠️ Build

Project: An automated evaluation harness for your RAG application:

A golden dataset of 30+ curated question-answer-ground-truth pairs.

Automated scoring for faithfulness, context precision, and answer relevance.

A GitHub Action CI workflow that halts deployment if quality metrics drop below baseline.

📁 Starter Code & Specification

✅ Done When

You can display benchmark scores comparing two different prompts or chunking configurations.
You identified and documented at least two edge cases where your LLM judge was unreliable or biased.
CI automatically runs evaluation checks on PRs.

Day 64–70: Tool Use and Agents

🎯 Focus

Empowering LLMs to interact with the external world: function calling, reasoning loops (ReAct), multi-step execution, step limits, and safety guardrails.

📖 Learn

Agent Architecture: Anthropic: Building Effective Agents
Agent Courses: Hugging Face Agents Course
Foundational Concepts: Microsoft: AI Agents for Beginners
Control & Safety: Tool schemas, human-in-the-loop approvals, infinite-loop breakers, and structured state machines.

🛠️ Build

Project: An autonomous CLI agent equipped with 3 distinct tools (e.g. web search, Python calculator, and local file reader). The agent maintains state, enforces a strict maximum step limit, and writes an audit log of every reasoning thought and tool execution.

📁 Starter Code & Specification

✅ Done When

The agent is provably protected against infinite loops via hard step boundaries.
You can document a specific scenario where a standard single-turn prompt was faster and cheaper than an agentic workflow.
Risky actions (like deleting a file or executing code) request human confirmation.

Day 71–77: MCP and AI Coding Workflows

🎯 Focus

The Model Context Protocol (MCP)—the emerging open standard for connecting AI models to external tools, databases, and enterprise data sources.

📖 Learn

Protocol Specifications: Model Context Protocol (MCP) Official Documentation
Architecture: MCP Hosts, Clients, and Servers; JSON-RPC communication.
Modern AI Workflows: How tools like Claude Code, Cursor, and Windsurf leverage MCP servers to interact with local development environments.

🛠️ Build

Project: A custom MCP server wrapping a real external API (e.g., GitHub, Slack, Linear, Weather, or your own database) with tested tools, schema validation, and integration with an MCP client (Claude Desktop or an agent).

📁 Starter Code & Specification

✅ Done When

An external MCP client can discover, inspect, and invoke tools on your server via STDIO or SSE.
You have written automated unit tests verifying tool inputs and outputs.
You can explain the architectural difference between custom function calling and the standardized MCP protocol.

Day 78–84: Deploy, Secure and Monitor

🎯 Focus

Taking AI systems to production: Docker containerization, security against prompt injection, LLM observability, cost tracking, and managed cloud deployments.

📖 Learn

Containerization: Docker: Get Started
LLM Security: OWASP Top 10 for LLM Applications
Observability: OpenTelemetry, Langfuse, or Arize Phoenix for trace logging.
Cloud Foundations: Amazon Bedrock Documentation and AWS Skill Builder Free Tier

🛠️ Build

Project: Containerize your FastAPI RAG / Agent service with a multi-stage Dockerfile, deploy it to a public endpoint (Render, Fly.io, or AWS), and configure request tracing, rate limiting, and a hard monthly budget kill-switch.

📁 Starter Code & Specification

✅ Done When

A live, public URL is accessible by anyone on the internet.
You executed prompt injection and jailbreak tests against your service and documented the defenses.
Request latencies, token counts, and costs are tracked in an observability dashboard.

Day 85–90: Capstone

🎯 Focus

Shipping a complete, portfolio-defining AI engineering product that ties together APIs, vector search/RAG, automated evals, tool integration, and cloud deployment.

🛠️ Build

Project: Build and ship an original, public AI product.

Your Capstone Repository must include:

Clear Value Proposition: A one-line hook and an animated demo GIF or video walkthrough.

Architecture Diagram: Mermaid or graphical diagram showing the data flow and system components.

Eval Scorecard: Benchmark accuracy, latency p95, and cost per request.

Post-Mortem & Learnings: What failed during development, design tradeoffs, and future roadmap.

📁 Capstone Guide & Template

✅ Done When

The repository is publicly hosted on GitHub with an immaculate README.
A stranger can sign up and test the live deployed demo in under 60 seconds.
Automated evals run inside GitHub Actions CI.

Bonus: Day 91–95: Portfolio, Resume and Applying

🎯 Focus

Translating your 90 days of engineering into career capital: portfolio curation, resume optimization, open-source contributions, and landing interviews.

📋 Action Checklist

GitHub Profile Curation: Pin your 2–3 best project repositories with clean thumbnails, descriptive badges, and clear README summaries.
AI Engineering Resume: Create a one-page technical resume emphasizing:
- Architecture decisions (RAG chunking strategies, hybrid search, rerankers).
- Measurable metrics (eval pass rates, latency reductions, cost caps).
- Links to live demos and GitHub CI badges.
Technical Write-up: Publish an in-depth article or blog post detailing a complex bug you solved, an eval harness you built, or a trade-off you navigated.
Open-Source Contribution: Submit a bug fix, documentation improvement, or feature PR to an open-source AI project you utilized during the 90 days.
Targeted Applications: Apply to high-conviction roles: AI Engineer, Applied AI Engineer, LLM Platform Engineer, or Backend Engineer (AI).

🔬 Optional Track: Under the Hood (Math and Deep Learning)

Do this track after completing Day 90, or alongside the main track if you want a foundational mathematical understanding of deep learning and model weights:

🧠 Zero to Hero Series: Karpathy: Neural Networks: Zero to Hero (build micrograd and nanoGPT from scratch).
📖 Interactive Textbook: Dive into Deep Learning (D2L.ai)
🎓 LLM Architecture Deep Dive: Maxime Labonne's Large Language Model Course
📐 Foundational Math:
- Linear Algebra: Matrix multiplication, eigenvectors, embeddings geometry.
- Calculus: Backpropagation, gradient descent, loss functions.
- Probability & Statistics: Maximum likelihood estimation, entropy, perplexity.

👨‍💻 Creator & Maintainer

Ahmed Raza Khan

AI Software Engineer | Cloud & AI Systems Integration

AI Engineer with 2+ years of production experience building and deploying AI-powered systems, intelligent agents, and automated tool workflows. Specializing in production LLM/NLP workflows (AWS Bedrock, OpenAI, Groq), RAG architectures, FastAPI microservices, and deep learning foundations in PyTorch.

Certifications: AWS Certified Machine Learning Associate • AWS Certified AI Practitioner • Claude Code Architect (CCA)

🌐 Website • 💼 LinkedIn • 🐙 GitHub • ✉️ Email

🌍 Translations

Help developers worldwide access this curriculum in their native languages!

Language	Status	Directory	Contributors
English	✅ Active (Source)	Root	Ahmed Raza Khan & Community
Urdu (اردو)	🚧 In Progress	`translations/urdu`	Contribute
Hindi (हिंदी)	🚧 In Progress	`translations/hindi`	Contribute
Spanish (Español)	🚧 In Progress	`translations/spanish`	Contribute
Arabic (العربية)	🚧 In Progress	`translations/arabic`	Contribute
French (Français)	🚧 In Progress	`translations/french`	Contribute

To contribute a translation, see our Translations Guide.

🤝 Contributing

Contributions make open-source thrive. Whether it's correcting a typo, updating a broken documentation link, proposing an improved starter project, or contributing a translation:

Read our Contribution Guidelines.
Check the Open Issues or start a Discussion.
Fork the repository, make your changes on a feature branch, and submit a PR!

Resource Addition Rule: Any proposed resource must be 100% free, up-to-date, and battle-tested by you personally.

👥 Community & Stargazers

If this roadmap helped you on your AI Engineering journey, please give us a Star ⭐—it helps more engineers discover this resource!

Built with ❤️ by Ahmed Raza Khan and the open-source community.

📜 License

This repository is distributed under the MIT License. See the LICENSE file for more information.

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
projects		projects
translations		translations
.gitignore		.gitignore
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
PROGRESS_TRACKER.md		PROGRESS_TRACKER.md
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

🚀 90 Days of AI Engineering

From "I can code a little" to shipping production AI products.

🌟 Why This Roadmap?

🗺️ Roadmap Architecture

📚 Table of Contents

📘 Introduction

👥 Who This Is For

🚀 Start Here

💰 Zero-Cost Guarantee

🛠️ Weekly Projects Overview

🗓️ Daily Curriculum

Day 1–14: Python, Git and APIs

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 15–21: How LLMs Work

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 22–28: Prompting and Structured Outputs

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 29–35: Backend with FastAPI

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 36–42: Embeddings and Vector Search

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 43–56: RAG (Two Weeks)

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 57–63: Evals

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 64–70: Tool Use and Agents

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 71–77: MCP and AI Coding Workflows

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 78–84: Deploy, Secure and Monitor

🎯 Focus

📖 Learn

🛠️ Build

✅ Done When

Day 85–90: Capstone

🎯 Focus

🛠️ Build

✅ Done When

Bonus: Day 91–95: Portfolio, Resume and Applying

🎯 Focus

📋 Action Checklist

🔬 Optional Track: Under the Hood (Math and Deep Learning)

👨‍💻 Creator & Maintainer

Ahmed Raza Khan

🌍 Translations

🤝 Contributing

👥 Community & Stargazers

📜 License

90DaysofAIEngineering