In the last six months I’ve been on hiring loops for AI engineer roles at three different stages of company, and I’ve seen the same thing every time.
A candidate walks in. They’ve finished four AI courses. They have a Github page of cloned notebooks. They can name fifteen models. They cannot tell me whether their last project ever ran outside the notebook it was written in.
This is the gap the title “AI engineer” is currently hiding.
Roughly seventy thousand people a month in India search the phrase “AI engineer” or close variants on Google. Most of them are looking for the same job. That job doesn’t exist. There are four jobs hiding under the title. Companies are hiring for one of them specifically. Knowing which one decides whether your prep gets you the role or wastes the next eighteen months.
Here’s what the four jobs actually are, what each one pays, and the honest path to get there.
The four jobs called “AI engineer”
In 2026, the title “AI engineer” maps to one of these four functions. Most job descriptions don’t say which one. You have to read the responsibilities section carefully and decode it.
One. The ML engineer. Trains and tunes models. Spends days on feature engineering, model selection, and validation. Works closely with data scientists. Knows their way around scikit-learn, PyTorch or TensorFlow, distributed training, and evaluation pipelines. Most of their week is in a notebook environment or in a training framework.
Two. The applied AI engineer. Doesn’t train models from scratch. Picks the right model for a real-world product feature, integrates it via an API, builds the surrounding software so it serves real users. Spends most of their week in regular product code (Python, Node, Go), with model calls as one component of a larger system.
Three. The data and ML platform engineer. Builds the infrastructure other AI engineers depend on. Feature stores, model registries, training pipelines, deployment systems, monitoring. Their week looks closer to a senior backend engineer than to a model trainer. Pay is typically the highest of the four because the talent pool is the thinnest.
Four. The LLM and agent engineer. Works specifically on systems built around large language models. Prompt engineering, retrieval-augmented generation, eval frameworks, tool-calling, agent orchestration. New role, hot job market in 2026, lots of titles, very few actual capable engineers.
When you see “AI engineer” on a job description, the responsibilities section usually maps to one of these four. The skill stack required and the salary band differ significantly across them.
The skill stack that actually gets hired
I run hiring loops. Across the four sub-specialties, here’s what consistently gets through versus what doesn’t.
Common across all four
Python depth. Not “I can write a script.” Comfortable with type hints, comfortable with the standard library, comfortable debugging through someone else’s code. Most candidates overestimate where they are. The interview catches it fast.
SQL fluency. Even AI engineers need to read data. If you cannot read a 100-line query and explain what it returns, you are slower than the rest of your team on the most common task at the job.
One deployment story. Have you put a model, or any service, into production. On what. Behind what. Behind what load balancer. Behind what monitoring. The candidates who clear interviews can describe at least one end-to-end deployment they were responsible for.
The debugger. Surprisingly diagnostic. Candidates who only know how to print-debug their code stand out. The ones who use a real debugger (pdb in Python, the equivalent in their IDE) tend to be the ones who can actually unblock themselves in production.
Reading a codebase you didn’t write. Three quarters of an engineer’s first year on the job is reading other people’s code. Most prep curriculum trains writing-from-scratch and skips reading. The bar in interviews has shifted accordingly.
Specialty-specific bars
ML engineer: Comfort with at least one deep-learning framework (PyTorch is the modal pick in Indian product companies in 2026). Hands-on experience with at least one of: distributed training, hyperparameter tuning at scale, or model compression. Has shipped at least one model whose predictions a real user has seen.
Applied AI engineer: Strong backend engineering. Comfort with at least one major model API (OpenAI, Anthropic, Google, an open-source one). Has built at least one product feature with a model in the request path that handles errors, latency, and cost.
ML platform engineer: Deep cloud engineering. AWS, Azure, or GCP, learned thoroughly. Comfort with Kubernetes, with model serving (Triton, BentoML, vLLM), with CI/CD pipelines for ML. The candidates I keep here are the ones who could equally have been hired as senior infra engineers.
LLM and agent engineer: Has built at least one retrieval-augmented system, ideally evaluated. Has wired up tool-calling end-to-end. Can name the failure modes of LLM agents and how their last project handled them. Bonus: has run an eval suite that compares two prompt variants quantitatively.
What doesn’t get you hired
Three certifications without one shipped project. The Coursera ML certificate that ten million other applicants also have. A portfolio of forked notebooks with no modifications. The phrase “I’ve worked with AI” without a single named system. A long list of acronyms (GPT, BERT, RAG, RLHF) with no specific story about where you’ve used them.
The candidates who get the call back have one strong specialty, one shipped artefact, and the ability to talk through their decisions without bluffing.
What a real AI engineer interview actually looks like
Walking into a 2026 AI engineer interview in India, expect roughly this shape across two to four rounds.
Round one is usually a recruiter screen. Quick. Resume sanity check, motivation, salary expectations, timeline. Most rejections at this stage are about misalignment on role or location, not capability.
Round two is typically a technical phone screen. One to two hours. The screener will want you to talk about a project you’ve shipped in real depth. They will ask about choices you made. They will ask what went wrong. Bluffing here is the most common failure mode. The candidate who says “I’m not sure but here’s how I’d find out” lands better than the candidate who fakes certainty.
Round three is the deep technical round. Two to four hours, sometimes split across days. Live coding (usually leetcode-medium with an ML twist or a system-design problem). ML fundamentals questions. Production scenarios. The bar is calibrated to the sub-specialty: ML engineer rounds dig into models, applied AI engineer rounds dig into product systems with model calls, ML platform rounds dig into infrastructure.
Round four, when it exists, is the leadership or culture round. The questions sound soft. They are not. The signal here is whether the candidate has worked under stress, made decisions with incomplete information, and learned from mistakes specifically and credibly.
A candidate who has thought about their own work in advance, can name two or three real decisions they made and why, and is comfortable saying “I don’t know yet” when they don’t, will clear most of these interviews. A candidate who has memorised a hundred algorithms but cannot talk about a single piece of work they own, will not.
AI engineer salary in India 2026
Honest ranges based on hiring loops I’ve seen in the last twelve months. Treat these as bands, not promises.
Fresher (0-1 year experience):
| Company type | Salary band |
|---|---|
| Service companies | ₹5-9 LPA |
| Product / GCC | ₹8-18 LPA |
| Top-tier product, well-funded startups | ₹18-30 LPA |
Mid-level (3-5 years):
| Company type | Salary band |
|---|---|
| Service companies | ₹12-22 LPA |
| Product / GCC | ₹22-45 LPA |
| Top-tier | ₹40-75 LPA |
Senior (6-10 years, owning production systems):
| Company type | Salary band |
|---|---|
| Service companies | ₹25-45 LPA |
| Product / GCC | ₹45-90 LPA |
| Top-tier and lead roles | ₹80 LPA - 1.5 Cr+ |
The gap between bands is not random. It tracks whether the company sells AI capability as a product, how much of your interview was about systems design versus tool recall, and whether you have a competing offer in hand. Sources: LinkedIn Salary, Glassdoor, and the GCC hiring patterns I’ve watched over the last 18 months.
The ML platform engineer band typically sits 15-25% above the others at the same experience level, because the role is the rarest. The LLM/agent engineer band is currently overpaying at the junior level (frothy market) and may correct.
Geography matters too, but less than candidates expect in 2026. Bangalore still leads salaries, with Hyderabad close behind. Pune, Chennai, and the Delhi NCR cluster sit 5-15% below Bangalore at equivalent levels. Remote-friendly roles have flattened the geography premium for senior candidates substantially in the last two years. For freshers, the office-location-based salary differential is still real.
How to become an AI engineer from a B.Tech base
Four stages. Each ends with a thing you should be able to do, not a course you have completed.
Stage 1, software engineering bones (months 1-3). Get your Python sharp. Get comfortable with one cloud’s basics. Ship one small product, end to end, that you would not be embarrassed to demo. This is the stage most candidates skip and most rejections trace back to.
Stage 2, pick a specialty (months 3-5). Read job descriptions for the four sub-specialties. Pick one. Stop bouncing between them. The candidate who has done three ML projects, one platform project, and two LLM projects looks unfocused to a hiring manager. The candidate with three solid LLM projects looks like a hire.
Stage 3, build the specialty-specific portfolio (months 5-12). For ML engineer, train a model end to end on a real dataset, deploy it, expose a working API. For applied AI engineer, build a real product feature using a model, with proper error handling and cost monitoring. For ML platform, deploy a serving system that someone else could use. For LLM and agent, build a retrieval system or an agent that actually works on something non-trivial, with evals.
Stage 4, interview, learn from rejections, repeat (months 12-18). Apply to 30 roles. Treat every interview as data. Note what you couldn’t answer. Go back and fix it. Repeat. Most candidates fail in the first ten interviews not because they were unprepared in absolute terms, but because they had no idea what specifically would be tested. After the first ten, the bar is clear, and the next twenty get progressively easier.
If you’re inside a programme that has industry projects from Year 2 onwards, you’re effectively running stages 1-3 inside the degree. If not, you’re doing this on weekends. Both work. The market doesn’t care which path got you there.
What the first 90 days on the job look like
For candidates who clear the interviews, the first 90 days at an AI engineer job often surprise them.
Week one is mostly setting up. Access requests. Reading documentation. Trying to understand how the team’s pipeline works. The instinct is to start writing code. The better instinct is to read code, ask questions, and map the system before adding to it.
Weeks two to four are usually small contributions. A bug fix. A small feature added to an existing pipeline. The code review process catches mistakes you didn’t realise you’d make. The candidates who do well here are not the ones who push back on feedback. They are the ones who treat feedback as a signal for what they don’t know yet.
Months two and three are when ownership starts to land. A small piece of the system becomes yours. You’re on-call for it. You wake up at 3 AM when it breaks. You realise that the production version of ML is fundamentally about uptime and trust, not about model accuracy. This is the moment that separates engineers who picked the field for the right reasons from those who didn’t.
The candidates who thrive past month three are the ones whose disposition matches the work: patient with infrastructure, persistent with bugs, comfortable with the slow build of a system other people can rely on.
The honest read on the market in 2026
Two things are true at the same time.
One. Demand for AI engineers in India is real, growing, and outpacing the talent supply for the next 3-5 years at minimum.
Two. The supply of people calling themselves AI engineers is growing even faster. Most of them are not the ones companies are hiring.
The way to be on the right side of that gap is not to take more courses. It’s to ship one real, end-to-end, production-quality piece of work in your chosen sub-specialty and to be able to talk about every decision in it.
The candidates who can do that get the offer. The ones who cannot do that don’t, regardless of how many certifications they list.
That’s the one thing worth taking from this piece. The rest is execution.
Anil is a co-founder of Kalvium and previously led engineering teams at Google and HackerRank. He runs hiring loops on a regular basis and writes about what the Indian tech market actually rewards. Read more from Anil or explore the AI-skills category.