A few months ago, a friend messaged me. Fresh out of a non-CS degree, wanted to get into AI. Had seen the headlines, felt like it was now or never. The question: "Where do I even start? I can't afford an M.Tech and I'm not sure a bootcamp is worth it."
I spent a weekend putting together a proper answer. Not a list of random YouTube links — an actual structured plan with specific resources, an honest timeline, and a real budget. This is that plan, cleaned up and published.
Everything in here comes from what I've personally used or seen work. I'm doing an M.Tech AI/ML at BITS Pilani right now and shipping production AI systems at HCLTech — so this isn't career advice from someone who read Medium posts. It's what I'd tell someone I actually care about.
How to read this: Each phase builds on the one before it. Do not skip ahead. Read the whole thing first, then start Phase 0 this week. The order matters more than the speed.
The honest truth about breaking in without a degree
Hiring managers at Indian AI companies — and increasingly at global ones — care about three things: GitHub projects, demonstrable skills, and cultural fit. A degree is a proxy for the first two. When you have actual projects, the proxy becomes irrelevant.
This is more true in AI than anywhere else. The field moves fast enough that a 2-year-old M.Tech curriculum is already behind. What matters is what you can build and show, right now.
The plan below takes 12 months at ~10 hours per week. Total maximum spend: ₹16,000. Most people reading this will spend under ₹4,000.
The 6 Phases
Math + Python Foundations
The non-negotiable bedrock. Do not skip or rush this.
Python Core Skills
- Variables, loops, functions, classes
- NumPy + Pandas for data
- Matplotlib + Seaborn for visualisation
- Git + GitHub from day one
- Jupyter Notebooks workflow
Math You Actually Need
- Linear Algebra — vectors, matrices, dot products
- Calculus — derivatives and chain rule (for backprop)
- Probability + Bayes' theorem
- Statistics — mean, variance, distributions
Resources: CS50P by Harvard (free on edX) — do every problem set. 3Blue1Brown "Essence of Linear Algebra" on YouTube (15 videos, ~5 hours). Khan Academy Statistics. All completely free.
Classical ML — The Scikit-learn Era
Core algorithms, proper evaluation, your first real project.
Algorithms to Master
- Linear + Logistic Regression
- Decision Trees + Random Forest
- SVM, KNN, Naive Bayes
- K-Means + PCA
- XGBoost / LightGBM
ML Engineering Foundations
- Train / validation / test splits — correctly
- Cross-validation + GridSearchCV
- Confusion matrix, ROC, AUC
- Feature engineering
- Handling imbalanced datasets
Milestone: Build one complete ML pipeline — raw data to evaluation — upload to GitHub with a README. Enter one Kaggle competition, even just a baseline submission. This one project is worth more than any certificate.
Deep Learning — PyTorch + Transformers
Neural networks, CNNs, RNNs, and the architecture that changed everything.
Neural Network Fundamentals
- Perceptrons → MLPs
- Backpropagation (understand it by hand)
- Activation functions, Dropout, BatchNorm
- CNNs for image classification
- RNNs and LSTMs for sequential data
The Transformer Era (2026 Must-Know)
- Attention mechanism — the single key insight
- BERT for NLP classification
- GPT architecture — how it actually works
- HuggingFace Transformers library
- Fine-tuning pre-trained models
Best resource: fast.ai Part 1 — free forever, no login required. Top-down approach: build first, understand theory after. It's the best deep learning course on the internet and Jeremy Howard is a legendary teacher. Also: d2l.ai (free book) and the HuggingFace course.
LLMs + Generative AI — The 2026 Layer
RAG pipelines, agentic systems, prompt engineering, fine-tuning. This is exactly what companies are hiring for.
LLM Engineering Skills
- Prompt engineering — zero-shot and few-shot
- Retrieval-Augmented Generation (RAG)
- Vector databases — FAISS, ChromaDB, Pinecone
- LangChain and LlamaIndex
- Fine-tuning with LoRA / QLoRA
- Evaluation using RAGAS
Agentic AI — The Hottest Skill in 2026
- Tool use and function calling
- Multi-agent orchestration
- Memory systems — short and long-term
- LangGraph, CrewAI, AutoGen
- Guardrails and hallucination handling
- Evaluating agent reliability
RAG pipelines and agentic systems are what "AI Engineer" job descriptions at Indian startups — Sarvam AI, Krutrim, Juspay, Groww — actually require in 2026. A working RAG chatbot on your GitHub will open more doors than any certification. Spend serious time here.
MLOps — Ship Models to Production
Docker, FastAPI, cloud platforms, CI/CD, monitoring. The skill that separates you from every other candidate.
Infrastructure and Deployment
- Docker — containerise your ML models
- FastAPI — serve models as REST APIs
- MLflow — experiment tracking + model registry
- GitHub Actions — automated CI/CD
- AWS SageMaker or GCP Vertex AI basics
Observability and Monitoring
- Prometheus + Grafana dashboards
- Model drift detection
- Data validation — Great Expectations
- A/B testing for ML model variants
- Cost optimisation for inference
Most ML candidates can train a model. Almost none can deploy one, monitor it, and handle it breaking in production. If you can train AND ship AND monitor end-to-end — you are in the top 20% of applicants.
Specialise + Build Portfolio + Apply
Pick one track, build 3 strong GitHub projects, start applying. Trying to master all four simultaneously is the most common mistake.
AI / ML Engineer
- LLMs + RAG + Agentic AI
- LangChain, HuggingFace
- Target: startups, product companies
MLOps Engineer
- Kubernetes + Kubeflow
- Model serving at scale
- Target: platform infrastructure teams
Your portfolio = 3 GitHub repos with good READMEs. Project 1: Classical ML pipeline on a real Kaggle dataset. Project 2: A working RAG chatbot with a REST API. Project 3: Something deployed and live — even on HuggingFace Spaces free tier with a working demo URL.
Your exact 12-month plan
Month 1–2 · Python + Math (₹0)
CS50P from Harvard — complete every single problem set. Kaggle Learn Python. Watch 3Blue1Brown "Essence of Linear Algebra" (15 videos, ~5 hours) and Khan Academy Statistics. Practical goal: write 50 lines of clean Python without Googling the syntax.
Month 3–4 · Classical ML on Kaggle (₹0 with audit)
Andrew Ng ML Specialization — audit free or apply for financial aid. Complete Kaggle "Intro to ML" and "Intermediate ML". Enter one Kaggle competition — even just a baseline submission counts. Upload that project to GitHub with a README.
Month 5–7 · Deep Learning with fast.ai (₹0)
fast.ai Part 1 — all 8 lessons, in order. Build an image classifier on a dataset you choose. Build a text sentiment classifier on movie reviews. Both go on GitHub with proper READMEs. These are your first two real portfolio projects.
Month 8–9 · LLMs + RAG + Agents (₹0)
HuggingFace NLP course — complete it fully. DeepLearning.AI short courses on LangChain and RAG (free during launch). Build a chatbot that answers questions over a PDF document. This is your star project that impresses every interviewer in 2026. Add it to GitHub with a live demo link.
Month 10–11 · Deploy + Cloud Basics (₹0–12k)
Wrap your best project in a FastAPI endpoint and Dockerize it. Deploy it on HuggingFace Spaces (completely free) so it has a live demo URL. Optionally start AWS Cloud Practitioner or Azure AI-900 prep if budget allows. Live deployed projects on a resume are genuinely rare and impressive.
Month 12 · Job Hunt Mode (₹0)
Polish LinkedIn with certifications and GitHub project links. Ensure each of your 3 repos has a complete README with screenshots and a demo link. Cold email 20 AI startups — most won't reply, but 2–3 will. Target: AI Engineer, ML Engineer, Junior Data Scientist at Indian product companies and funded AI startups.
The exact courses — free first
These are in the order you should take them. Each one assumes vocabulary and hands-on skill from the previous. Skipping ahead is the most common mistake — resist it.
- CS50P — Harvard Introduction to Python (edX, free, 8 weeks) — The best free Python course, period. Harvard certificate carries genuine weight. Do every problem set.
- Kaggle Learn — ML + Python Micro-courses (free, 2–4 weeks) — Zero setup, runs in browser, instant feedback. Take in order: Intro to ML → Intermediate ML → Pandas → Data Visualisation.
- fast.ai — Practical Deep Learning for Coders (free forever, 10 weeks) — Top-down approach: build first, understand theory after. Used by working ML researchers. No login, no paywall, no time limit.
- HuggingFace NLP Course (free, 6 weeks) — Non-negotiable in 2026. Over 90% of AI job descriptions mention HuggingFace directly. Maps to real daily work at AI companies.
- DeepLearning.AI Short Courses (free at launch, 1–3 hours each) — Take: ChatGPT Prompt Engineering, Building Systems with the ChatGPT API, LangChain for LLM Development, Building and Evaluating Advanced RAG Applications.
Courses worth paying for
- Andrew Ng ML Specialization — Coursera (₹3–4k or free via financial aid) — The most recognised ML cert globally. 3 courses: Supervised ML, Unsupervised ML + Recommenders, Intro to Reinforcement Learning. Apply for financial aid — approval rate is very high and you get 100% off.
- AWS Cloud Practitioner or Azure AI Fundamentals (₹8–12k exam fee) — Take this at Month 10 only — not before. Cloud certs are powerful but only meaningful after you have solid ML fundamentals. Before that, it's theory with no context.
Exactly how much will this cost?
| CS50P — Harvard Introduction to Python | ₹0 — completely free |
| Kaggle Learn — all ML and Python courses | ₹0 — completely free |
| fast.ai — Practical Deep Learning Part 1 | ₹0 — free forever |
| HuggingFace NLP Course | ₹0 — completely free |
| DeepLearning.AI short courses (LangChain, RAG) | ₹0 — free at launch |
| Andrew Ng ML Specialization (financial aid) | ₹0–4,000 |
| Cloud certification exam — AWS / Azure (optional) | ₹8,000–12,000 |
| HuggingFace Spaces — host and deploy projects live | ₹0 — free tier is sufficient |
| Total maximum spend over 12 months | ₹12,000–16,000 |
Honest truths nobody tells you
GitHub beats certificates every time
3 solid GitHub projects with proper READMEs beats 10 certificates on a resume. Recruiters at Indian AI companies check GitHub before they read anything else. No GitHub = effectively invisible to hiring managers.
Build something every two weeks
Passive tutorial-watching doesn't build skills — writing actual code does. Set a rule: every two weeks, you ship something to GitHub. Even a single Jupyter notebook with analysis counts. The habit of shipping matters.
Pick ONE specialisation and go deep
AI Engineer or MLOps Engineer or Data Scientist — pick one and go all in. Trying to be great at all three simultaneously means being great at none. In 2026, employers actively prefer specialists over shallow generalists.
LLMs + RAG = job offers in 2026
A working RAG chatbot deployed on HuggingFace Spaces is worth more than an ML theory certificate. This is exactly what funded AI startups — Sarvam AI, Krutrim, Juspay — are actively hiring for right now.
One more thing
I sent my friend a version of this guide. Six months later, they had two GitHub projects live, had finished fast.ai, and were deep into building their first RAG chatbot. They hadn't spent a single rupee yet.
The resources are all there. The structure is there. The only thing between you and a job offer is whether you start Phase 0 this week or next month. Start this week.
If you're already past some of these phases and want to discuss specific architecture decisions, RAG implementations, or MLOps setups — I'm happy to talk. Use the contact page.