The 2026 AI Roadmap for Indian Engineering Students

In FY26, AI-skilled graduates made up 60% of TCS’s fresher hires, up from 10 to 15% three years ago. That number comes from TCS CHRO Sudeep Kunnumal at the AI Impact Summit in March 2026 (Rediff/Business Standard). The roadmap below is for the engineering student who wants to be on the right side of that statistic by their final-year placement window.

A roadmap article is only useful if it cuts through the noise. So before the curriculum, three honest framings: not every job in 2026 needs AI (the standard service-tier roles still hire on aptitude plus basic coding). The free curriculum is genuinely as good as anything paid. And two deployed projects on a public GitHub beat any stack of certificates. We’ll cover all three with specifics below.

What “AI engineer” actually means in Indian hiring, 2026

The title “AI engineer” only became common in Indian job postings around 2024. Before that, almost everything in this space was called either “ML engineer” or “data scientist”. The titles still overlap, and recruiters use them interchangeably more often than they should. Here’s the working frame most hiring managers use in 2026.

Role	What they actually do	Math depth	Engineering depth	Typical fresher CTC band
Data Scientist	Analyse data, build statistical models, communicate findings to business	High (statistics, A/B testing)	Lower (notebooks, less production code)	₹4.5 to 12 LPA
ML Engineer	Train, evaluate, and deploy ML models in production	Medium (linear algebra, probability)	High (Python, MLOps, infra)	₹6 to 15 LPA
AI Engineer (the 2026 title)	Build applications on top of existing LLMs and AI APIs (RAG, agents, LLM orchestration)	Lower (you don’t train the base model)	High (Python, system design, evals)	₹6 to 15 LPA
Applied Research Scientist	Push the state of the art (rare for freshers, mostly hires from research labs)	Very high	Medium	₹15 to 35 LPA (rare; under 200 hires/year nationally)

Three things follow from this table. First, the bulk of new fresher hiring in 2026 is for the AI Engineer role, not the older ML Engineer or Data Scientist titles. The reason is structural: most companies are now consumers of AI, not producers. They use OpenAI, Anthropic, and open-weight models from Hugging Face; they don’t train their own from scratch. The roadmap below is calibrated for this AI Engineer role specifically, with the ML Engineer track as a natural next step.

Second, the math you need is less than you think for the application-layer roles, and more than you think for the model-training roles. If your math comfort is low, head for AI Engineer first; you can deepen the math later if you want to move toward ML Engineering.

Third, “Applied Research Scientist” is genuinely rare at fresher level in India. Don’t optimise your roadmap for it unless you already have a strong publication record from undergrad research.

Where AI fits into your placement timeline

This is the section to read carefully if you’re a final-year engineering student staring at both placements and AI prep, and not sure how to fit them together. The blunt answer: AI prep does not replace placement prep, and it doesn’t compete with it. They run in parallel for different drives.

The drives at your campus split into roughly three categories.

Service-tier mass hiring (TCS Ninja, Infosys System Engineer, Wipro Standard, Capgemini, HCL TechBee). Starting CTC ₹3.5 to 4.5 LPA, with industry-wide moderation in fresher hiring across this tier in FY26 (Financial Express). Selection is aptitude plus basic coding plus communication. AI prep doesn’t help much here, and skipping aptitude to do AI is a mistake. If your goal is “any IT job by graduation”, the aptitude work in our TCS placement preparation guide and the spoken-English work in the AMCAT SVAR cornerstone is what gets you offers. AI is a nice-to-have at this tier.

Higher-tier service hiring (TCS Prime, TCS Digital, Infosys Specialist Programmer, Infosys Power Programmer, Wipro WILP+). Starting CTC ₹6 to 11 LPA. Selection is aptitude plus stronger coding plus a clear technical screen. AI prep starts to matter at this tier, especially for TCS Prime: the TCS CHRO Sudeep Kunnumal statement that 60% of new joiners are AI-skilled isn’t marketing language, it’s a screen criterion. The Infosys Power Programmer track similarly favours candidates with demonstrable AI project work, beyond their HackWithInfy ranking.

Product and startup hiring (Atlassian, Razorpay, PhonePe, Postman, Freshworks, Zoho, plus a long tail of Series A and B startups). Starting CTC varies widely; the salary table later in this article shows the full breakdown. Selection is technical depth, system design basics, and project portfolio. AI is now table stakes at this tier; without an AI project on your GitHub, your resume often doesn’t pass the first filter.

The practical sequencing for a final-year student with both placements and an AI roadmap to chase: do aptitude prep on weekday evenings (sustainable, low-cognitive-load); use weekends for the AI roadmap (high-focus, deep-work blocks). The math works out to roughly 6 hours a week of aptitude plus 10 hours a week of AI work, which most students can sustain through a 4-month placement window without burning out. The total weekly cognitive load is comparable to one well-run elective course.

If you’re at Wipro’s campus drive or one of their 50 university Centres of Excellence, AI exposure now gets you into a different (and better) interview track. The CoE selection process is one of the cleanest examples in 2026 of AI prep paying off inside an existing IT-services hiring funnel.

The realistic salary picture

Salary expectations for fresher AI roles in India are wildly distorted by clickbait articles citing fifty-lakh fresher offers. Those exist. They go to fewer than 200 candidates nationally per year, almost all from IIT and a handful of the top NITs and BITS, with publication records or ICPC finals on their resumes. For everyone else, the realistic 2026 picture looks like this.

Company tier	Typical fresher CTC (AI-track roles)	What they want to see
IT services baseline (TCS Ninja, Infosys SE, Wipro Standard)	₹3.5 to 4.5 LPA	Aptitude plus basic coding; AI is bonus
IT services AI-tier (TCS Prime, Infosys Specialist/Power Programmer, Wipro CoE)	₹6.5 to 11 LPA	Aptitude plus 1 deployed AI project plus stronger coding screen
IT services elite AI-tier (HCLTech 2026 program for AI-skilled freshers)	₹18 to 22 LPA	Strong AI / GenAI / data engineering portfolio plus extended technical screen
Mid-size IT and GCCs (Tech Mahindra, Mphasis, ZS Associates, GCC analytics teams)	₹5 to 9 LPA	1 to 2 AI projects, strong Python, basic SQL and data work
Product companies entry (Razorpay, Freshworks, Postman, Zoho)	₹8 to 15 LPA	2+ deployed projects, system design basics, strong fundamentals
Funded AI-first startups (Series A and B)	₹8 to 18 LPA	2+ deployed projects with non-trivial complexity, willingness to take ownership
Top product companies (Atlassian, Adobe, PhonePe, Microsoft IDC)	₹15 to 28 LPA	Strong fundamentals plus DSA plus 2 substantial projects plus system design
FAANG-tier (Google, Meta, Amazon AWS, Microsoft Research, Anthropic, OpenAI)	₹25 to 45+ LPA	Top-tier college plus publications or competitive programming achievements; under 200 fresher offers/year nationally

The honest read: if you’re at a Tier-2 or Tier-3 college without a research record, the realistic aspirational target for your first AI job is the “Product companies entry” band shown in the table above, at a product company or AI-first startup or GCC. Aiming for the FAANG-tier band as a baseline plan is setting yourself up to feel like you failed when you land at the upper end of the IT services AI-tier band, which is in fact a good outcome by industry data.

A separate read: the IT services AI-tier band shown above is the most under-discussed track in 2026. TCS Prime, Infosys Power Programmer, and Wipro CoE-selected hires earn 2 to 3 times what their service-tier peers earn at the same company, with the same degree, often from the same campus drive (TCS CHRO interview, Rediff/Business Standard). The differentiator is one or two deployed AI projects on a public portfolio. That ROI is hard to beat.

The ceiling case in IT services is HCLTech’s 2026 elite program for AI-skilled freshers, with offers in the ₹18 to 22 LPA range (The Hans India). It’s the highest publicly disclosed IT-services AI premium for entry-level hires, and signals where the rest of the sector’s AI-tier compensation is heading. The screen for this band is a strong AI portfolio plus an extended technical interview, not a separate competitive exam, which makes it more reachable for a serious Tier-2 college student than the FAANG track.

The six-stage roadmap

This is the curriculum, sequenced. Each stage names the free resource that’s actually best, the time budget at 10 hours a week, and what you should be able to do at the end. We’re picking one resource per stage rather than a generic list, because the most common mistake students make is collecting tabs and never finishing any one course.

Stage 1: Python, math foundations, and Git (4 to 6 weeks)

Goal: Be able to write a clean Python script that reads data, does basic numerical work, and is version-controlled.

Resource: If you already know one programming language, Python’s official tutorial plus Git’s pro book is enough; you don’t need a video course. If Python is new, the University of Helsinki’s MOOC.fi Python Programming course is free and excellent.

Math: You need working comfort with vectors and matrices (linear algebra), probability and basic statistics, and the chain rule from calculus. 3Blue1Brown’s Essence of Linear Algebra YouTube series teaches the visual intuition in about 4 hours. Pair it with Khan Academy’s statistics and probability course for the gaps.

You’re done with Stage 1 when: you can clone a GitHub repo, set up a Python virtual environment, install dependencies, run someone else’s notebook, push your changes back, and explain to a friend what a derivative does intuitively.

Stage 2: ML foundations (3 to 4 weeks)

Goal: Understand what machine learning is at the level of a working practitioner, not a researcher. Build and evaluate your first models.

Resource: Andrew Ng’s Machine Learning Specialization on Coursera (the new 2022 version, not the old 2011 Stanford one). You can audit it free. It covers linear regression, logistic regression, neural networks, decision trees, and the overfitting/regularisation framework you’ll use in every interview question for the rest of your career.

You’re done with Stage 2 when: you can take a tabular dataset (anything from Kaggle), build a model that predicts something useful, evaluate it honestly using train/validation/test splits, and explain why your model isn’t working when it isn’t.

Stage 3: Deep learning (4 to 6 weeks)

Goal: Understand neural networks well enough to read papers, debug models, and explain to a hiring manager what’s actually happening when you train one.

Resource: fast.ai’s Practical Deep Learning for Coders, free, 9 lessons. The course teaches by doing: by lesson 2 you’ve trained and deployed a working model. The fast.ai approach (top-down, code-first, theory-as-needed) is genuinely the right way to learn this material in 2026, and the free version is identical to what paid bootcamps teach for ₹50,000.

Optional but excellent companion: Andrej Karpathy’s Neural Networks: Zero to Hero. Free, six videos, builds GPT from scratch in code. If you want to be the kind of AI engineer who can read a transformer paper and follow what’s happening, watch all six. The Karpathy material is harder than fast.ai but rewards every hour you put in.

You’re done with Stage 3 when: you can train a CNN on an image classification task, train a basic recurrent model on text, fine-tune a pretrained model from Hugging Face on your own data, and articulate what backpropagation is doing under the hood.

Stage 4: LLMs, RAG, and AI applications (3 to 4 weeks)

Goal: Build the kind of AI applications that the bulk of “AI engineer” jobs in 2026 are actually hiring for.

Resource: The Hugging Face Course, free, covers Transformers, fine-tuning, datasets, tokenizers, and the LLM application patterns that ship in production. Pair it with LangChain’s Get Started guide and the Pinecone learning centre for vector databases.

The pattern you’re learning at this stage is: load a pretrained LLM (or call an API), add retrieval over your own data (RAG), wrap it in a Python service, and deploy it. This is what you’ll do on day 1 of an AI engineer job in 2026.

You’re done with Stage 4 when: you can build a question-answering bot over a custom document collection (your college’s syllabus PDFs, say), make it run with a real LLM API, get the retrieval to actually return relevant chunks, and ship the whole thing as a working web app.

Stage 5: Deployment, evals, and production basics (2 to 3 weeks)

Goal: Move from “it works on my laptop” to “it runs reliably for someone else.”

Resource: FastAPI’s official tutorial for building Python APIs, Docker’s getting started for containerisation, and any free Vercel or Render tutorial for deployment. For evals, the Anthropic prompt engineering documentation is a clean entry point to thinking about model output quality.

This stage is short on content and long on practice. You’ll spend most of it actually deploying things, not reading.

You’re done with Stage 5 when: the project from Stage 4 is running at a public URL, has a basic eval suite that catches regressions, and you can explain to a stranger what’s running where and why.

Stage 6: A portfolio project (4 to 6 weeks)

Goal: Build one substantial project that’s actually useful to someone, deploy it, and put it on your resume.

Resource: None. This is your work. Pick a problem that’s specific to you (your college, your hobby, a process at home, a friend’s small business) and build the AI tool you wish existed. Deploy it. Make a 2-minute video walkthrough. Write a one-page README on the GitHub repo explaining what you built and why.

The four projects in the next section give you templates if you don’t know where to start.

You’re done with Stage 6 when: your project has at least one user who’s not you, a GitHub repo with clean commit history, and a README that a recruiter can read in 90 seconds and understand what you built.

Projects that change a recruiter’s mind

Most resumes show three projects with one-line descriptions (“Built an ML model for sentiment analysis using Python and TensorFlow”). Most recruiters skip past these in 4 seconds. The projects below are the kind that actually get clicked.

Project 1: Document Q&A bot for your college

Take your college’s syllabus PDFs, exam timetables, placement policy documents, fee structure, and library catalogue. Build a chatbot that answers questions over them with citations. RAG with a free embedding model (sentence-transformers), a vector DB (ChromaDB local or Pinecone free tier), and either OpenAI’s API or a local Llama model.

Why recruiters click: It’s specific (named college, real documents), shows the full RAG pipeline, and is genuinely useful. Several students have turned this project into a paid contract with their college after graduating.

Resume line: “Built and deployed a RAG-based chatbot over [college name]‘s academic and placement documents (over ten thousand text chunks, ChromaDB, OpenAI API, FastAPI backend, Vercel deploy). Used by 80+ students for fee, syllabus, and placement queries.”

Project 2: Resume parser and matcher

Build a tool that takes a job description and a resume PDF, scores the fit, and explains the gaps. Use a small LLM for the parsing and matching, with structured output (JSON Mode if using OpenAI, or function calling).

Why recruiters click: It’s a recruiter-relevant problem (they understand the use case immediately), and it shows you can work with structured outputs, which is a standard 2026 production pattern.

Resume line: “Built a resume-to-JD matching tool using LLM-based parsing and structured outputs. Matches across 17 dimensions (skills, experience, education, projects). Deployed at [URL]; processes ~200 resumes/week from college users.”

Project 3: A small AI agent

Build an agent that does one specific task end-to-end: a calendar manager that responds to natural-language requests and creates Google Calendar events, a personal finance bot that reads SMS bank alerts and categorises spending, or a hobby-specific tool (a chess move analyser, a recipe scaler that handles substitutions).

Why recruiters click: Agentic AI is the buzzword of 2026 hiring. A working agent demonstrates that you understand the loop (tool use, planning, error recovery) without needing the candidate to talk through it.

Resume line: “Built a multi-step AI agent for [task] using LangGraph; handles tool use across 4 APIs (Calendar, Email, Notion, Slack); recovers from API failures and logs decisions for debugging.”

Project 4: A fine-tuned model for a niche task

Pick a domain where general-purpose models are bad (Tamil literature classification, mechanical part identification from images, a regional cuisine recipe generator) and fine-tune a small open-weight model on it. Hugging Face’s free tools and a Google Colab GPU are enough.

Why recruiters click: Fine-tuning is the skill that separates “AI user” from “AI engineer” in many hiring rubrics. Even a small fine-tune with a 200-row dataset shows you understand the workflow.

Resume line: “Fine-tuned a 1B-parameter open-weight model on a custom ~2k-example dataset for [task]. Improved task accuracy by roughly 35 to 40 percentage points over the base model (full evaluation table on the model card). Public model and dataset on Hugging Face; weights downloaded 200+ times.”

A note on the resume lines above: every one ends with a number (users, accuracy, downloads, weekly volume). That’s deliberate. AI projects without numbers attached read as homework; with numbers, they read as professional work, even when the numbers are small. Recruiters scan for numbers; help them.

The mistakes most students make

Five recurring patterns we see in students working through an AI roadmap. Each one comes with a specific fix.

Mistake 1: Collecting courses, finishing none. The most common pattern. A student starts fast.ai, then sees Karpathy is excellent and switches, then hears about a new Hugging Face course and starts that. None get finished. Pick one course per stage, finish it, then move on. The next course will still be there in 4 weeks.

Mistake 2: Skipping the math. Tempting because the math feels boring next to LLM tutorials. But interviewers test it: “explain backpropagation”, “what’s the difference between L1 and L2 regularisation”, “why does batch normalisation work”. The 8 hours of 3Blue1Brown videos save you days of awkward interviews.

Mistake 3: Building from scratch when you should be using libraries. The opposite mistake also exists, but for freshers in 2026 the bigger problem is reinventing the wheel. You don’t need to build a tokeniser from scratch (use Hugging Face’s). You don’t need to write your own vector store (use ChromaDB). The hiring signal is whether you can compose existing tools into a working system, not whether you can implement transformers in NumPy. Build the agent, ship it, and learn the internals only when a specific problem demands it.

Mistake 4: Not deploying anything. A model in a Jupyter notebook is invisible to a recruiter. The same model behind a public URL with a 30-second screen recording is a portfolio piece. Free deployment options (Vercel, Render, Hugging Face Spaces) exist specifically for this. The Stage 5 work is the difference between “studied AI” and “built AI”.

Mistake 5: Buying an expensive paid course you won’t finish. Two patterns here. One: paid courses cost more than the value they add over the free curriculum (with the cohort-accountability and academic-cert exceptions noted above). Two: paid courses, especially the forty-to-fifty-thousand-rupee tier, often produce a perverse motivation effect. Students who pay big often feel they’ve done the hard part by paying, and then under-invest in actually doing the work. The free curriculum has the property of demanding more from you, which is good for learning. Spend the same money on a laptop that runs a local LLM, or on a cohort-based program if accountability is your bottleneck. Don’t spend it on yet another video course.

A note on what comes next

Once you’ve worked through Stages 1 to 4, you have two natural next steps depending on where you’re aiming.

If you want to start building immediately and learn the LLM application patterns hands-on (RAG, agents, structured outputs, evals), the fastest entry point is a self-paced playground. TinkerLLM is built for exactly this transition: a low-friction environment to write your first production-grade LLM code without configuring infrastructure first. It’s the bridge between “watched a course” and “shipped something useful”.

It’s a higher commitment than the free curriculum, but for students who learn better with structure and accountability, it’s the cleanest version of the roadmap above.

Both paths route from the same starting point: this article. Pick the one that matches your learning style, not the one that sounds more impressive.