Mathematical & statistical foundations
Linear algebra, probability, statistics, and the calculus that underwrites modern machine learning. Not as an exam paper — as the language a student uses to reason about why a model behaves the way it does.
Reads and explains vector, matrix, and basic probability notation; Read full ↓ Derives gradients for standard loss functions (cross-entropy, MSE); Read full ↓ Selects appropriate statistical tests for an experimental design they wrote themselves; Read full ↓ Derives novel loss functions or regularisers from first principles for a domain-specific problem; Read full ↓ Reads contemporary theoretical ML literature in their sub-area and integrates it into shipped work; Read full ↓ Implementation skill
Writing, debugging, profiling, and maintaining the code that trains and serves AI systems. Measured against working software engineers, not against student exercises.
Reads existing PyTorch or JAX code; Read full ↓ Writes a new model class (e. Read full ↓ Authors a small AI service end-to-end (model, API, container, basic observability); Read full ↓ Builds custom training infrastructure where off-the-shelf tooling falls short (custom samplers, mixed-precision training, multi-GPU sharding); Read full ↓ Authors original training infrastructure including custom kernels (Triton, CUDA) where required; Read full ↓ System design
Composing pipelines — retrieval-augmented generation, agents, evaluation harnesses, data pipelines — and choosing the right model and data flow for the actual problem in front of you.
Diagrams the high-level components of a standard AI system (a chatbot, a classifier, a basic RAG pipeline); Read full ↓ Builds a working single-purpose AI system end-to-end (e. Read full ↓ Designs a multi-component AI system from a requirements document; Read full ↓ Designs novel architectures for new problem types — agentic loops, multi-stage retrieval, human-in-the-loop pipelines — including the evaluation and rollback story; Read full ↓ Sets system-design standards for an organisation; Read full ↓ Evaluation & methodology
Designing benchmarks, ablations, error analysis, and the statistical rigour that lets a team know whether a change actually helped. The dimension most often missing from self-taught practitioners.
Computes standard accuracy, precision, recall, and F1; Read full ↓ Designs an evaluation set appropriate to a small project, including held-out data; Read full ↓ Designs a domain-specific evaluation harness from scratch; Read full ↓ Designs evaluation methodology for novel system types where no standard benchmark exists (agentic systems, generative outputs, retrieval-grounded answers); Read full ↓ Defines evaluation standards for an organisation or sub-field; Read full ↓ Safety, alignment & responsible AI
Bias evaluation, red-teaming, regulatory mapping (EU AI Act, India's DPDP Act), model and system cards. Not a compliance checkbox — a working part of how a serious AI engineer ships.
Names the major categories of AI harm (bias, privacy, hallucination, misuse) and gives a concrete example of each; Read full ↓ Performs a basic disaggregated evaluation across an obvious sensitive attribute; Read full ↓ Designs a bias-evaluation methodology suitable to the project's actual deployment context; Read full ↓ Designs the responsible-AI process for a team — incident response, post-deployment monitoring, escalation paths; Read full ↓ Sets responsible-AI policy at organisation level; Read full ↓ Industry communication
Explaining technical decisions to non-technical buyers, writing model cards a lawyer can read, and defending design trade-offs to senior reviewers. The dimension Indian technical education most reliably under-trains.
Explains what a given AI system does and does not do, in plain English, to a non-technical reader; Read full ↓ Writes a project design note a non-technical product manager can act on; Read full ↓ Defends a system design to a sceptical senior reviewer (engineering manager, professor, industry mentor) including trade-offs not made; Read full ↓ Writes the technical narrative for a product launch or a research publication; Read full ↓ Recognised externally as a credible technical voice in their sub-area — invited talks, cited posts, op-eds in serious publications; Read full ↓ 