AI / ML Engineer
You become the person who makes machines learn, and proves they learned the right thing.
Behind every 'AI feature' is someone who chose the model, trained or fine-tuned it, measured it honestly and kept it fast in production. That person understands what's inside the black box. Your maths degree is the unfair advantage here: gradients, probability and linear algebra are the actual working tools.
Overnight training finished. Validation loss looks great, which is suspicious. You find data leakage: duplicates across train and test.
Fixed the split, retrained. Accuracy is lower but honest, and it now matches what happens in production.
Product asks: 'Fine-tune or just prompt GPT?' You run both on 200 examples. The fine-tuned 1B model matches it at 1/30th the cost.
You quantise the model, cut latency from 400 ms to 90 ms, and ship it behind an API.
The drift dashboard is green. You queue tomorrow's experiment and go home.
What you'll build
- Models that classify, rank, predict and generate
- Fine-tuned LLMs for specific tasks
- Search and recommendation systems
- Evaluation suites and ML pipelines in production
Your maths edge
This is the most maths-heavy role on the list. Backprop is the chain rule, embeddings are linear algebra, loss functions come from probability, and evaluation is statistics. Most people learn these as black boxes. You'll know what they're doing.
Demand
AI/ML specialists are among the WEF's fastest-growing jobs to 2030. 64% of new GCC roles in India in 2026 needed AI or data skills.
Entry pay (India, rough)
Roughly ₹6–18 LPA in India, with a long tail far above that at top product companies and AI labs.
The honest truth
Pure 'research' roles need an MS or PhD. Applied ML engineering, which is building and shipping models, is open to strong BTech graduates with real projects. Kaggle-only portfolios aren't enough. Show deployed, measured systems.
Try a few minutes of the job
Three real moments from an ML engineer's week. Choose, then see how an experienced engineer reasons. There's no score and no right path for you. It's just a feel for the work.
Your fraud model has 99.2% accuracy. Your manager is thrilled. Only 0.5% of transactions are fraud. Thoughts?
From your first line of code to people using what you built
Tick off steps as you go. Finishing a quest reveals a sealed insight, and finishing a stage unlocks a capability. Everything is saved in this browser.
Before you start: where to use AI (and where not to)
Every quest has its own "use AI for / avoid" note. The full guide is in the AI Playbook.
Sprint Zero
Five days. Python, the maths you'll actually use, and the core libraries. You'll go deeper while building.
Learn by building
Eight builds from scratch to state of the art. Each one makes the next make sense.
Company-style builds
Real ML briefs: an ambiguous problem, a business goal, and a definition of done. Choosing the approach is up to you.
Architect mode
You know how the pieces work, so stop typing every line. From here on you design, write the spec, direct AI agents and review their work like a tech lead.
Your meaningful project
One project that real people use beats ten tutorial clones. This becomes the story in every interview.