The 12-Week Study Plan
Eight to twelve hours a week, four phases, and a concrete deliverable every week — because a working artefact beats a completed reading list in every deep-dive round. Halve the pace for a 24-week part-time schedule, or compress to six weeks at 20+ hours.
Plan progress
0 of 12 weeks complete
Phase 1 — Foundations
Rebuild the maths and classical ML that every screen assumes.
- 8–10 hrs
Week 1
Linear algebra & probability
Be able to derive the backprop gradient of a linear layer and explain cosine similarity from first principles.
Deliverable: Implement PCA from scratch and match sklearn to 1e-8.
- 8–10 hrs
Week 2
Statistics & experimentation
Size an A/B test in your head and explain why offline AUC gains vanish online.
Deliverable: Build the experiment analysis toolkit (sample size, SRM, bootstrap CIs).
- 10–12 hrs
Week 3
Classical ML fundamentals
Diagnose overfitting from learning curves and name all five categories of data leakage.
Deliverable: Leakage-proof evaluation harness with grouped + time-based splits.
- 10–12 hrs
Week 4
Trees, metrics & feature engineering
Explain why boosting overfits, why feature_importances_ misleads, and derive a cost-optimal threshold.
Decision Trees, Random Forests & Gradient…Evaluation Metrics: Choosing and Defendin…Feature Engineering, Imbalanced Data & Un…Deliverable: Tuned GBM baseline with SHAP analysis and a calibration report.
Phase 2 — Modelling Depth
Deep learning and LLM internals you can derive on a whiteboard.
- 10–12 hrs
Week 5
Neural networks & backpropagation
Derive backprop through softmax + cross-entropy from memory, with correct shapes.
Deliverable: NumPy neural-net library with finite-difference gradient checks.
- 8–10 hrs
Week 6
Training, optimisation & regularisation
Explain AdamW, warmup, and how you would debug a NaN loss at step 12,000.
Deliverable: Instrumented training run with per-layer gradient norms and an induced divergence.
- 12–14 hrs
Week 7
Transformers & attention
Derive scaled dot-product attention, justify √d_k, and size a KV cache.
Deliverable: GPT-style transformer with RoPE, GQA and a working KV cache.
- 10–12 hrs
Week 8
LLM fundamentals & fine-tuning
Explain tokenisation cost, Chinchilla, LoRA maths, and RLHF vs DPO.
Deliverable: LoRA fine-tune with a data-mixing sweep and a general-benchmark check.
Phase 3 — Production & Systems
RAG, agents, MLOps and the ML system design framework.
- 12–14 hrs
Week 9
LLM inference, RAG architecture
Explain PagedAttention throughput maths and design a RAG pipeline end to end.
Deliverable: RAG system with hybrid retrieval, reranking and a 100-query gold set.
- 12–14 hrs
Week 10
Agents & MLOps
Design a bounded agent with guardrails; explain training/serving skew and drift detection.
AI Agents, Tool Calling & Production Guar…Model Deployment: Serving Patterns, Featu…Monitoring, Drift Detection & Incident Re…Deliverable: Serving endpoint with a shared feature module, data contracts and behavioural tests.
- 10–12 hrs
Week 11
ML system design
Run a full 45-minute design out loud using the 8-step framework, twice.
Deliverable: Three written system designs with back-of-envelope arithmetic.
Phase 4 — Interview Polish
Mock loops, story bank, take-home practice, offer strategy.
- 12–15 hrs
Week 12
Coding drills, stories & offer strategy
Implement k-means and attention under a timer; deliver three defensible project stories.
NumPy, pandas & ML Algorithms From ScratchPyTorch Fluency & Debugging Under Intervi…Project Deep-Dives & Behavioural Rounds f…Interview Strategy, Take-Homes & Negotiat…Deliverable: Story bank, one full mock loop, and a comparison model for offers.
Five things that make the difference
Spaced repetition beats rereading
Use the flashcard drill for 10 minutes daily rather than rereading lessons. Retrieval practice is what moves knowledge into recall you can use under pressure.
Speak your answers out loud
An answer you can write is not an answer you can say. Record yourself on three questions a week — the gap between the two is usually large and always fixable.
Build one thing per phase
The challenge at the end of each lesson exists so you have something concrete to talk about. A working artefact beats a completed reading list in every deep-dive round.
Start mock interviews in week 6, not week 12
Mocks surface communication gaps that solo study cannot. Six weeks of feedback beats two weeks of panic, and early mocks tell you which topics to prioritise.
Apply to a warm-up company first
Your first loop is always the worst. Schedule a company you like but are not desperate for, treat it as calibration, and go into your target loops already warmed up.
Ready for week 1?
Start with linear algebra. It underpins embeddings, attention and every gradient you will be asked to derive.
Begin lesson 1Curriculum
- Machine Learning Fundamentals: The Screening Round5 Q · 18 min
- Linear & Logistic Regression Deep Dive5 Q · 19 min
- Decision Trees, Random Forests & Gradient Boosting5 Q · 20 min
- Evaluation Metrics: Choosing and Defending the Right One5 Q · 18 min
- Feature Engineering, Imbalanced Data & Unsupervised Learning5 Q · 19 min