Month-by-Month Execution Plan¶

M1 through M13 — The Complete Arc¶

This document is the operational truth of the 13-month plan. Every month has a theme, concrete deliverables, exit checks, and alignment notes for your Zoho context. Read your current month at the start of every sprint. Read the next month at the end of every second sprint.

The Valley of Despair is Months 5-7. You will feel like you are behind, that the math is too hard, that deep learning is impenetrable, that everyone else is faster. This is not a sign of failure. It is a sign you are learning hard things. It is named here in advance so that when you arrive, you recognize it and continue anyway.

Buffer weeks: Week 8, Week 16, Week 22, Week 26. These are yours. Use them to catch up, rest, or consolidate — never to add new material.

Month 1 — The Foundation Is Poured¶

Theme: Linear algebra and calculus become your first language. Phase: Phase 0 (primary) Sprints: S1, S2 Weekly hours: 10-12 (calibration phase — find your actual pace)

Deliverables:

Complete 01_phase_0_mathematical_foundations/01_linear_algebra.md study + all exercises
Complete 01_phase_0_mathematical_foundations/02_calculus_and_optimization.md study + exercises
Lab notebook started: 13_discipline/02_lab_notebook.md template in use by Week 2
First entry in your prediction log (see 07_phase_6_research_and_mastery/04_developing_technical_opinion.md)

Learning targets:

Gilbert Strang OCW 18.06 Lectures 1-14 (through eigenvalues)
3Blue1Brown Essence of Linear Algebra — all 15 videos
Khan Academy Multivariable Calculus for calculus intuition

Portfolio rung status: Rung 1 in progress (SVD + gradient descent notebook)

Zoho alignment: Quiet month. Observe: which ML feature at Zoho do you interact with daily? Write one sentence about what might be powering it. No action required yet.

India-specific notes: July-August heat. Keep study space cool or early-morning sessions before 8am. Monsoon disruptions to power/internet — download resources locally in Week 1.

Exit check (must pass before Month 2):

Can you compute a matrix inverse by hand (2×2)?
Can you explain why eigendecomposition matters for PCA?
Can you derive the chain rule and apply it to a 3-layer function composition?
Lab notebook has at least 4 entries.
Gradient descent implemented in pure NumPy, tested on a quadratic function.

Month 2 — The Mathematical Foundation Completes¶

Theme: Probability becomes a tool, not a symbol system. Phase: Phase 0 (primary) → Phase 1 begins (last week) Sprints: S3, S4 (S4 = Buffer Week 8 at end) Weekly hours: 12-14

Deliverables:

Complete 01_phase_0_mathematical_foundations/03_probability_and_statistics.md + exercises
Complete 01_phase_0_mathematical_foundations/04_information_theory.md
Rung 1 SHIPPED: Math From Scratch notebook — SVD implementation + gradient descent with visualization, published to GitHub
Phase 0 exit criteria fully checked

Learning targets:

Blitzstein & Hwang “Introduction to Probability” Chapters 1-7
3Blue1Brown probability videos
Begin 02_phase_1_classical_ml/01_supervised_learning.md

Portfolio rung status: Rung 1 complete ✅ | Rung 2 begins

Zoho alignment: Notice where Bayesian reasoning or A/B testing language appears in your team’s work. Write one paragraph in your lab notebook connecting it to what you just learned. Still observing, not proposing.

Exit check:

Can you derive Bayes’ theorem from first principles?
Can you explain KL divergence geometrically?
Rung 1 notebook is public on GitHub with README.
Buffer week (Week 8) used for rest or review, not new material.

Month 3 — Classical ML: The Backbone¶

Theme: Supervised learning — the engine of 90% of deployed ML. Phase: Phase 1 (primary) Sprints: S5, S6 Weekly hours: 12-15

Deliverables:

Complete 02_phase_1_classical_ml/01_supervised_learning.md + all code exercises
Complete 02_phase_1_classical_ml/02_unsupervised_learning.md
sklearn installed, working, used on at least 2 real datasets (not toy)
First Kaggle account created, first dataset downloaded

Learning targets:

sklearn MOOC (official) — Modules 1-4 (this is genuinely the underrated gem; 574 upvotes for a reason)
Hands-On ML with Scikit-Learn (Géron) Chapters 1-6
Andrew Ng ML Specialization (2022 Python version) as secondary reference

Portfolio rung status: Rung 2 in progress (Classical ML Battle)

Zoho alignment: Identify one classification or regression problem at Zoho that has existing data. Don’t build anything yet. Just identify it and describe the feature space in your lab notebook (3 sentences).

Valley of Despair early warning: If sklearn feels “too easy,” you are probably using it correctly at the surface and will hit the hard parts in Month 4. Do not skip to deep learning early. The classical foundation is not optional.

Exit check:

Implemented logistic regression from scratch (gradient descent, no sklearn).
Can explain bias-variance tradeoff with a specific example and numbers.
t-SNE failure modes: can you name all 5? (They’re in 02_phase_1_classical_ml/02_unsupervised_learning.md)

Month 4 — Classical ML: Pipeline Mastery¶

Theme: From “I can fit a model” to “I can build a production-grade pipeline.” Phase: Phase 1 (primary) → Phase 2 begins (last week) Sprints: S7, S8 (S8 = Buffer Week 16) Weekly hours: 12-15

Deliverables:

Complete 02_phase_1_classical_ml/03_model_selection_and_evaluation.md
Complete 02_phase_1_classical_ml/04_feature_engineering.md
Complete 02_phase_1_classical_ml/05_classical_ml_in_production.md
Rung 2 SHIPPED: Classical ML Battle — 3-model comparison on a public dataset, full evaluation (ROC, PR curve, calibration), GitHub repo with reproducible pipeline

Learning targets:

sklearn MOOC Modules 5-7
Hands-On ML Chapters 7-9 (ensembles, pipelines)
Fast start on 03_phase_2_deep_learning_core/01_neural_networks_from_scratch.md

Portfolio rung status: Rung 2 complete ✅ | Rung 3 begins

Zoho alignment: You now have enough vocabulary to have one informal conversation with a senior at Zoho about the ML feature you identified in Month 3. Ask what model they’re using. Ask how they evaluate it. Listen. Write notes. Don’t propose anything yet.

Exit check:

5-fold cross-validation implemented and understood mechanistically (why 5 and not 10?).
Rung 2 repo is public, README is clear, pipeline is reproducible from scratch.
Buffer week (Week 16) used wisely — no new topics.

Month 5 — ⚠️ VALLEY OF DESPAIR BEGINS — Deep Learning Core¶

Theme: Neural networks from scratch. This is where most people stop. You will not stop. Phase: Phase 2 (primary) Sprints: S9, S10 Weekly hours: 13-15

Deliverables:

Complete 03_phase_2_deep_learning_core/01_neural_networks_from_scratch.md
Complete 03_phase_2_deep_learning_core/02_pytorch_fundamentals.md
Backpropagation implemented in pure NumPy, tested on XOR problem
PyTorch working on your local machine or cloud GPU

Learning targets:

Karpathy’s micrograd from scratch (YouTube, 2 hours — mandatory)
Fast.ai Part 1 Chapters 1-4 (top-down intuition to complement bottom-up)
cs231n Stanford (2024 version) Lectures 1-7
3Blue1Brown Neural Networks series (all 4 episodes)

Portfolio rung status: Rung 3 in progress (Deep Learning From Scratch)

Zoho alignment: GPU compute question becomes real. Check 11_tools_setup/02_gpu_and_compute_strategy.md for INR-priced options. Decide on your cloud strategy this month.

Valley of Despair note: Backpropagation will feel like it doesn’t make sense for 3-5 days. This is normal. Karpathy’s micrograd video is the cure. Watch it. Implement it. Watch it again. The fog will lift.

Exit check:

Backprop implemented from scratch — correct gradients verified against PyTorch autograd.
Training loop written in PyTorch: forward → loss → backward → optimizer step.
You can explain what a computation graph is without looking it up.

Month 6 — ⚠️ VALLEY OF DESPAIR CONTINUES — CNNs and Sequences¶

Theme: Vision and sequences. The two domains that made deep learning indispensable. Phase: Phase 2 (primary) → Phase 3 begins (last week) Sprints: S11, S12 (S12 = Buffer Week 22) Weekly hours: 13-15

Deliverables:

Complete 03_phase_2_deep_learning_core/03_cnns_and_computer_vision.md
Complete 03_phase_2_deep_learning_core/04_sequence_models.md
Complete 03_phase_2_deep_learning_core/05_training_dynamics.md
Rung 3 SHIPPED: Deep Learning From Scratch — full backprop engine, CNN on CIFAR-10, CharLM with training curves, all on GitHub

Learning targets:

cs231n Lectures 8-14 (CNNs, RNNs, attention preview)
Karpathy’s makemore series (character language model from scratch)
PyTorch Lightning introduction (for training boilerplate)

Portfolio rung status: Rung 3 complete ✅ | Rung 4 begins

Zoho alignment: You now have enough depth to have a genuine technical conversation about any DL-based feature at Zoho. Request a 30-minute coffee chat with someone working on a DL problem. Goal: understand their stack, not sell yourself.

Valley of Despair reality check: By end of Month 6, you will have been at this for 6 months with a full-time job. You will feel tired and possibly behind. This is the exit of the Valley. Phase 3 (Transformers) is where things start to click. The investment you made in math and classical ML is about to compound.

Exit check:

CNN trained on CIFAR-10, >85% test accuracy.
Rung 3 all three artifacts public.
model.eval() vs torch.no_grad() — can explain the difference and the bug they prevent.
Buffer week (Week 22) taken. Non-negotiable.

Month 7 — The Architecture Revolution¶

Theme: Transformers. The architecture that changed everything. Phase: Phase 3 (primary) Sprints: S13, S14 Weekly hours: 13-15

Deliverables:

Complete 04_phase_3_modern_architectures/01_attention_and_transformers.md
Complete 04_phase_3_modern_architectures/02_bert_and_gpt_family.md
Self-attention implemented from scratch with correct tensor shapes
“Attention Is All You Need” (Vaswani 2017) read and annotated

Learning targets:

Karpathy’s nanoGPT video (3 hours — the best transformer tutorial in existence)
Andrej Karpathy’s GPT from scratch implementation — follow line by line
“The Illustrated Transformer” (Jay Alammar) — archived, still canonical for intuition
HuggingFace NLP course Chapters 1-4

Portfolio rung status: Rung 4 in progress (Transformer Lab)

Zoho alignment: LLM-based features at Zoho are likely production by now. Ask: what’s their inference stack? What model? What latency SLA? You’re building context for Q3 proposal (Month 9).

Exit check:

Can derive scaled dot-product attention including the √d_k normalization rationale.
nanoGPT trained on a small corpus, loss curve descending.
Can explain why BERT is not autoregressive and why GPT is.

Month 8 — Architectures: Vision, Diffusion, Graphs¶

Theme: The full scope of modern deep learning beyond NLP. Phase: Phase 3 (primary) → Phase 4 begins (last week) Sprints: S15, S16 Weekly hours: 13-15

Deliverables:

Complete 04_phase_3_modern_architectures/03_vision_transformers.md
Complete 04_phase_3_modern_architectures/04_diffusion_models_foundations.md
Complete 04_phase_3_modern_architectures/05_graph_neural_networks.md
Rung 4 SHIPPED: nanoGPT + domain fine-tune + DDPM MNIST (all 3 artifacts public)

Learning targets:

ViT paper (Dosovitskiy 2020) — read original
“Denoising Diffusion Probabilistic Models” (Ho 2020) — implement the DDPM in the fortress file
Cora GCN implementation using PyTorch Geometric

Portfolio rung status: Rung 4 complete ✅ | Rung 5 begins

Zoho alignment: You now have a complete mental model of the architecture landscape. By end of Month 8, write a 1-page internal document (for your own records, not to share yet) about where Zoho’s ML stack sits relative to SOTA. This is preparation for Month 9’s proposal.

Exit check:

DDPM trained on MNIST — 8×8 sample grid generated, loss curve correct.
Can explain over-smoothing in GNNs as a theorem (not an observation).
Rung 4 all artifacts public.

Month 9 — ⛔ HARD GATE APPROACHING — Generative AI Frontier Begins¶

Theme: RLHF, DPO, PEFT. The alignment and efficiency layer of modern LLMs. Phase: Phase 4 (primary) Sprints: S17, S18 Weekly hours: 14-15

Deliverables:

Complete 05_phase_4_generative_ai_frontier/01_llm_alignment_and_finetuning.md
Complete 05_phase_4_generative_ai_frontier/02_parameter_efficient_finetuning.md
QLoRA fine-tune attempt on a small model (Llama 3.2 1B or similar) — even if rough
Hard Gate 1 planning: define your target fine-tune task for Month 10

Learning targets:

“Training language models to follow instructions with human feedback” (InstructGPT, OpenAI 2022)
“Direct Preference Optimization” (Rafailov 2023) — read the math section twice
“QLoRA: Efficient Finetuning of Quantized LLMs” (Dettmers 2023)
HuggingFace PEFT library documentation

Portfolio rung status: Rung 5 critical path — hard gate is M10

Zoho alignment: Q3 starts. This is the month for one concrete proposal to your manager about applying LLM fine-tuning or RAG to a specific Zoho problem. Keep it small. Show the math. Bring a baseline. See 00_command/06_background_alignment.md for the script.

Exit check:

Can derive DPO loss from the RLHF constrained optimization (the reparameterization trick).
QLoRA fine-tune running, even if results aren’t polished yet.
Have a specific target task for Rung 5.

Month 10 — ⛔ HARD GATE 1: LLM Engineering¶

Theme: You either ship the LLM engineering gate or you don’t. There is no partial credit. Phase: Phase 4 (primary) Sprints: S19, S20 Weekly hours: 14-15 (this is the most intensive month)

Deliverables:

Complete 05_phase_4_generative_ai_frontier/03_rag_systems.md + implement full RAG pipeline
Complete 05_phase_4_generative_ai_frontier/04_llm_agents.md
⛔ RUNG 5 SHIPPED — ALL 4 ARTIFACTS REQUIRED:
1. Fine-tuned model weights on HuggingFace Hub (with model card)
2. RAGAS evaluation scores in a public document (not screenshots — a script)
3. Live accessible demo URL (HuggingFace Spaces or similar)
4. RAG pipeline code on GitHub

Learning targets:

“RAGAS: Automated Evaluation of RAG” (Es 2023)
LangChain/LlamaIndex documentation (pick one, master it)
“Toolformer: Language Models Can Teach Themselves to Use Tools” (Schick 2023)

Portfolio rung status: Rung 5 complete ✅ (if gate passes) | Rung 6 begins

Gate failure protocol: If any of the 4 Rung 5 artifacts does not exist by end of Month 10, do not proceed to Phase 5. Use the reset protocol in 99_pre_mortem/09_summary_and_reset_protocol.md. Fix Rung 5 first.

Exit check:

All 4 Rung 5 artifacts are public and accessible to an interviewer clicking a link.
RAGAS scores are real numbers from a real evaluation run, not placeholders.
Demo URL loads and returns an inference result within 30 seconds.

Month 11 — Production: The Real World¶

Theme: Serving, monitoring, experiment tracking. The gap between “model works” and “model ships.” Phase: Phase 5 (primary) Sprints: S21, S22 Weekly hours: 12-14 (slight deceleration — consolidation phase)

Deliverables:

Complete 06_phase_5_production_and_mlops/01_experiment_tracking_and_versioning.md
Complete 06_phase_5_production_and_mlops/02_model_serving_and_inference.md
MLflow running locally, tracking experiments from any prior rung
vLLM deployment of a small LLM, latency benchmarked

Learning targets:

MLflow and W&B official documentation (not tutorials — docs)
vLLM paper: “Efficient Memory Management for Large Language Model Serving with PagedAttention”
“Reliable Machine Learning” (Google SRE for ML) — relevant chapters

Portfolio rung status: Rung 6 critical path — hard gate is M12

Zoho alignment: Q4 begins. If Month 9’s proposal was accepted, you are now building something real. If not, that’s fine — your external portfolio is more important than internal alignment at this stage.

Exit check:

MLflow tracking server running, at least 20 experiment runs logged.
vLLM serving a 1B model, throughput measured vs. HuggingFace baseline.
Can explain PagedAttention in 2 minutes without slides.

Month 12 — ⛔ HARD GATE 2: Production ML System¶

Theme: Ship a live, monitored, CI/CD-backed ML system. Not a demo. A system. Phase: Phase 5 (primary) + Phase 6 begins (last week) Sprints: S23, S24 Weekly hours: 14-15 (final intense push)

Deliverables:

Complete 06_phase_5_production_and_mlops/03_containerization_and_orchestration.md
Complete 06_phase_5_production_and_mlops/04_monitoring_and_data_quality.md
Complete 06_phase_5_production_and_mlops/05_ci_cd_for_ml.md
⛔ RUNG 6 SHIPPED — ALL 5 ARTIFACTS REQUIRED:
1. Live running inference endpoint (not localhost)
2. DVC + MLflow experiment lineage visible in repo
3. GitHub Actions CI/CD with green runs in history
4. Evidently monitoring report showing drift detection on real data
5. Runbook document (written, not theoretical)
Rung 7 SHIPPED: Paper reproduction — one ICLR/NeurIPS/ICML 2023-2025 paper, code on GitHub, deviation analysis documented

Portfolio rung status: Rungs 6 and 7 complete ✅ | Rung 8 begins

Gate failure protocol: Same as Rung 5. Do not enter Month 13 mastery phase with an incomplete production system. Fix it first.

Exit check:

System is live at interview time — verifiable with a URL.
GitHub Actions history shows at least 5 successful pipeline runs.
Evidently report has been run against real production data, not test fixtures.
Runbook exists and would enable someone else to restart the system from scratch.

Month 13 — The Applied PhD Identity¶

Theme: Research fluency, open source, technical opinion. Becoming someone the field recognizes. Phase: Phase 6 (primary) Sprints: S25, S26 (S26 = Buffer Week 52) Weekly hours: 10-12 (deliberate deceleration — consolidation and projection)

Deliverables:

Complete all files in 07_phase_6_research_and_mastery/
Rung 8 SHIPPED: 3 technical posts published (Substack/Medium) + one conference talk submitted or one paper review submitted
First open source PR merged (documentation counts — 85% merge probability)
M13 pitch rehearsed in 4 versions: 30-sec, 2-min, 5-min technical, portfolio landing

Learning targets:

Read 10 papers from 07_phase_6_research_and_mastery/02_frontier_papers_2024_2025.md
Run the prediction log review from 07_phase_6_research_and_mastery/04_developing_technical_opinion.md
Review job market: 07_phase_6_research_and_mastery/06_interview_and_assessment_mastery.md

Portfolio rung status: All 8 rungs complete ✅

Zoho alignment: If you want to use this work for internal positioning — you are ready. If you want to use it for external opportunities — you are ready. Both are valid. What you are not is trapped.

Closing note: Month 13 is not an ending. It is a new baseline. The person who finishes this fortress can read any ML paper and understand the claims, implement any architecture from scratch, debug any production failure, and explain the entire stack to anyone. That’s not a credential. That’s capability. Build from here.

Exit check (Month 13 = Final Audit):

All 8 portfolio rungs are public and accessible.
M13 pitch sentence can be delivered under 45 seconds, cold.
Lab notebook has entries for at least 40 of the 52 weeks.
At least one person in your network knows about this journey.

Return to: README.md · Next: 02_sprint_calendar.md