2025 Edition · Comprehensive Depth Roadmap

ArtificialIntelligence& Machine Learning

From linear algebra and Python to deep learning, transformers, MLOps, and building production AI systems — a rigorous, honest roadmap that doesn't skip the hard parts.

Start learning Track your progress

0Modules

0–18Months

0+Topics

0Projects

PythonLinear AlgebraStatisticsScikit-learnPyTorchTensorFlowTransformersLLMsComputer VisionNLPMLOpsReinforcement Learning

// prerequisites

Mathematical Foundations

This is what separates engineers who use ML from engineers who understand ML. Don't skip this. Study it in parallel with early modules.

Linear Algebra

Vectors, matrices, matrix multiplication, dot products, eigenvalues/eigenvectors, SVD, PCA. Everything in ML is linear algebra under the hood.

Statistics & Probability

Distributions (Gaussian, Bernoulli, Poisson), Bayes' theorem, MLE, MAP estimation, hypothesis testing, p-values, confidence intervals, CLT.

Calculus & Optimization

Derivatives, partial derivatives, chain rule, gradients, gradient descent, Hessians, Lagrange multipliers. Backprop is just the chain rule.

Information Theory

Entropy, cross-entropy, KL divergence, mutual information. These appear constantly in loss functions and model evaluation.

Numerical Methods

Floating point arithmetic, numerical stability, condition numbers. Understand why nan/inf appear and how to prevent them.

Graph Theory (Basics)

Nodes, edges, adjacency matrices, graph traversal. Essential for GNNs and understanding attention mechanisms as fully-connected graphs.

Key mathematical objects you'll use daily

Loss: L(θ) = -1/N · Σ [y·log(ŷ) + (1-y)·log(1-ŷ)]

Gradient: ∇θL = ∂L/∂θ

Update: θ ← θ - α · ∇θL

Attention: Attn(Q,K,V) = softmax(QKᵀ/√dₖ) · V

Bayes: P(θ|X) ∝ P(X|θ) · P(θ)

// learning_path

10-Module Deep Roadmap

Each module builds on the last. Core theory → implementation → project. Click any module to reveal hidden depth.

core topics

Python fluency: list comprehensions, generators, decorators, context managers
NumPy: ndarray, broadcasting, vectorized ops, einsum — the ML workhorse
Pandas: DataFrames, groupby, merge, apply, time-series indexing
Matplotlib + Seaborn: plotting distributions, correlations, model results
Jupyter notebooks: workflow, magic commands, widgets
SciPy: optimization, signal processing, stats
Virtual environments: venv, conda, poetry
Type hints, dataclasses, pydantic for clean ML code

⬡ milestone

You can load a CSV, clean messy data, run statistical summaries, visualize distributions and correlations, and write vectorized NumPy operations that process 1M rows in milliseconds — no loops.

Project: Exploratory Data Analysis Pipeline

Load a real dataset (Kaggle/UCI). Full cleaning, outlier detection, feature distributions, correlation heatmaps, and a written EDA report.

essential libraries

numpypandasmatplotlibseabornscipyplotlypolarspydantic

bonus depth

Polars vs Pandas: when to switch for 10x speed on large data
Dask: parallel computing for datasets that don't fit in RAM
Profiling Python code: cProfile, line_profiler, memory_profiler
C extensions and Numba for ultra-fast numerical code

supervised learning

Linear Regression: OLS, Ridge (L2), Lasso (L1), ElasticNet — derive the normal equation
Logistic Regression: sigmoid, cross-entropy loss, multi-class (softmax)
Decision Trees: ID3/CART algorithm, Gini impurity vs entropy, pruning
Random Forests: bagging, feature importance, OOB error
Gradient Boosting: XGBoost, LightGBM, CatBoost — the Kaggle weapons
SVM: hyperplane, kernel trick (RBF, polynomial), soft margin
k-NN: curse of dimensionality, distance metrics, approximate NN
Naive Bayes: generative model, MAP classification

unsupervised learning

K-Means, K-Medoids: initialization (K-Means++), convergence, elbow method
DBSCAN: density-based clustering, epsilon, min_samples
PCA: variance explained, scree plot, reconstruction
t-SNE, UMAP: high-dim visualization
Gaussian Mixture Models: EM algorithm

model evaluation (critical)

Bias-Variance tradeoff — the most important concept in ML
Cross-validation: k-fold, stratified, time-series CV
Classification metrics: precision, recall, F1, AUC-ROC, PR-AUC
Regression metrics: MAE, MSE, RMSE, R², MAPE
Confusion matrix, calibration curves, SHAP values
Hyperparameter tuning: GridSearch, RandomSearch, Optuna (Bayesian)
Feature engineering: encoding, binning, interactions, scaling
Imbalanced data: SMOTE, class weights, threshold tuning

⬡ milestone

You can implement linear regression from scratch using gradient descent, explain why gradient boosting usually wins on tabular data, select the right metric for a given business problem, and tune a model with Optuna.

Project: Tabular Prediction Competition

Solve a Kaggle tabular competition end-to-end: EDA, feature engineering, baseline → XGBoost → LightGBM ensemble. Beat baseline by >10%. Write a full model card.

tools

scikit-learnxgboostlightgbmcatboostoptunashapimbalanced-learnfeature-engine

bonus depth

Implement gradient descent for linear regression in pure NumPy
AutoML: AutoGluon, H2O.ai — understand what they do under the hood
Anomaly detection: Isolation Forest, LOF, Autoencoders
Survival analysis: Kaplan-Meier, Cox regression

from scratch first

Perceptron → MLP: forward pass, activation functions, output layer
Backpropagation: derive it fully from the chain rule — not optional
Activations: ReLU, Leaky ReLU, GELU, Swish, Sigmoid, Tanh — when to use each
Loss functions: MSE, BCE, CrossEntropy, Focal, Huber
Weight initialization: Xavier/Glorot, He, and why it matters
Vanishing/exploding gradients: causes and solutions

training techniques

Optimizers: SGD, Momentum, Adam, AdamW, Lion — intuition and math
Learning rate schedules: cosine annealing, warmup, OneCycleLR
Regularization: Dropout, L1/L2, BatchNorm, LayerNorm, WeightDecay
Batch size effects: large batch training, gradient accumulation
Early stopping, model checkpointing, training curves
Mixed precision training: fp16, bf16, torch.autocast

pytorch (go deep)

Tensors: creation, ops, broadcasting, GPU (.cuda() / .to(device))
Autograd: computational graph, .grad, retain_graph, no_grad
nn.Module: forward(), parameters(), buffers, hooks
DataLoader + Dataset: custom dataset, collate_fn, sampler
Training loop best practices: gradient clipping, AMP, DDP
torch.compile (PyTorch 2.x) for 2–3x speed
TensorBoard / W&B for experiment tracking

⬡ milestone

You can implement a neural network in pure NumPy (forward + backprop), then reproduce it in PyTorch and verify gradients match. You can train a model on GPU with mixed precision and log experiments to W&B.

Project: Neural Net from Scratch + PyTorch Reproduction

Build a 3-layer MLP in NumPy (with backprop). Reproduce in PyTorch. Train on MNIST/CIFAR-10. Add BatchNorm, Dropout, LR scheduling. Compare all optimizers. Log to W&B.

tools

PyTorchTensorFlow/KerasWeights & BiasesTensorBoardLightningeinopstorchmetrics

bonus depth

Implement Adam optimizer from scratch in NumPy
Custom autograd: extend torch.autograd.Function for custom ops
JAX + Flax: functional approach to deep learning
torch.compile deep dive: Dynamo, TorchInductor

convolutional networks

Convolution operation: filters, stride, padding, dilation, groups
Pooling: max, average, global average pooling (GAP)
Architectures: LeNet → AlexNet → VGG → ResNet (skip connections!) → EfficientNet
Transfer learning: feature extraction vs fine-tuning vs linear probing
Data augmentation: RandomCrop, ColorJitter, Mixup, CutMix, AugMix
Image preprocessing: normalization, ImageNet stats, PIL vs OpenCV

object detection

R-CNN → Fast R-CNN → Faster R-CNN: region proposals, RoI pooling
YOLO (v5, v8): single-shot detection, anchor boxes, NMS
SSD, RetinaNet, DETR
mAP metric: IoU thresholds, PR curve per class
Semantic vs instance vs panoptic segmentation
SAM (Segment Anything): zero-shot segmentation

vision transformers

ViT: patch embeddings, positional encoding, class token
DeiT, Swin Transformer, ConvNeXt
CLIP: contrastive image-text pretraining — multimodal foundation
DINO, DINOv2: self-supervised visual features
Diffusion models for images: DDPM, stable diffusion architecture

⬡ milestone

You can fine-tune a ResNet50 on a custom dataset, train a YOLO model for object detection, achieve >90% accuracy on a 10-class problem, and implement a custom augmentation pipeline in Albumentations.

Project: Custom Object Detector

Collect/annotate data with Label Studio. Fine-tune YOLOv8. Evaluate with mAP. Export to ONNX. Deploy as a REST API with FastAPI.

tools

torchvisiontimmAlbumentationsultralytics (YOLO)OpenCVLabel StudioONNXDetectron2

bonus depth

Video understanding: 3D convolutions, optical flow, temporal models
Medical imaging: DICOM, 3D segmentation with nnU-Net
Generative models: GANs (StyleGAN), VAEs, Stable Diffusion fine-tuning (LoRA, DreamBooth)
Neural Radiance Fields (NeRF) for 3D scene representation

classical NLP

Text preprocessing: tokenization, lemmatization, stemming, stopwords
Bag of Words, TF-IDF, n-grams
Word2Vec, GloVe, FastText — static word embeddings
RNN, LSTM, GRU — sequential models and their problems
Seq2Seq with attention: the precursor to Transformers

transformers (deep)

Attention mechanism: queries, keys, values — derive from scratch
Multi-head attention, positional encoding (sinusoidal + RoPE)
Transformer architecture: encoder, decoder, encoder-decoder
BERT: masked LM, NSP, fine-tuning paradigm
GPT: autoregressive LM, causal attention mask
T5, BART: text-to-text and denoising
Tokenization deep dive: BPE, WordPiece, SentencePiece, tiktoken
Hugging Face: pipeline, AutoModel, Trainer API, datasets

NLP tasks

Text classification, sentiment analysis, NER, POS tagging
Question answering: extractive (SQuAD) and generative
Text generation: beam search, sampling, top-k, top-p, temperature
Summarization: extractive vs abstractive
Machine translation evaluation: BLEU, ROUGE, BERTScore
Semantic similarity: sentence-transformers, cosine similarity, FAISS

⬡ milestone

You can implement the attention mechanism from scratch, fine-tune BERT for classification, build a semantic search system with sentence-transformers + FAISS, and generate text with a decoder-only model.

Project: Semantic Search Engine

Embed a document corpus with sentence-transformers. Build a FAISS vector index. Query interface with FastAPI. Fine-tune a classifier on domain-specific data. Evaluate with proper metrics.

tools

transformerstokenizerssentence-transformersFAISSspaCyNLTKdatasets (HF)evaluateaccelerate

bonus depth

Implement multi-head attention from scratch in PyTorch
Flash Attention 2: memory-efficient attention for long contexts
Efficient transformers: Longformer, BigBird, linear attention
Cross-lingual models: mBERT, XLM-R, multilingual embeddings

LLM architecture & training

Scaling laws: Chinchilla, compute-optimal training
Modern LLM architecture: RoPE, GQA, SwiGLU, RMSNorm, KV cache
Distributed training: data parallelism, tensor parallelism, pipeline parallelism (3D parallelism)
Training data: curation, deduplication, filtering (FineWeb, Dolma)
Instruction tuning: SFT on curated chat data
RLHF: reward model, PPO, Constitutional AI
DPO, ORPO, SimPO: preference optimization without RL

fine-tuning techniques

LoRA: low-rank adaptation — math intuition (W = W₀ + BA)
QLoRA: 4-bit quantization + LoRA for single-GPU fine-tuning
Prompt tuning, prefix tuning, adapter layers
Full fine-tuning vs PEFT trade-offs
Dataset formats: Alpaca, ShareGPT, ChatML
Evaluation: MMLU, HellaSwag, GSM8K, MT-Bench, custom evals

prompting & inference

Prompt engineering: zero-shot, few-shot, chain-of-thought, tree-of-thought
RAG: chunking, embedding, retrieval, reranking, generation
Quantization: GPTQ, AWQ, GGUF (llama.cpp)
Inference optimization: speculative decoding, continuous batching, PagedAttention (vLLM)
Tool calling / function calling: ReAct, structured outputs
Guardrails: output safety, factuality, hallucination detection
LLM agents: planning, memory, tool use (LangGraph, smolagents)

⬡ milestone

You can fine-tune Llama 3 on a custom task with QLoRA, evaluate it against a baseline, build a RAG system with reranking, and serve it with vLLM with streaming responses.

Project: Domain-Specific LLM + RAG System

Fine-tune Llama 3.1 8B on domain data with QLoRA. Build a RAG pipeline (LangChain + ChromaDB). Evaluate with RAGAS. Deploy with vLLM. Build a Gradio UI.

tools

transformerspefttrlbitsandbytesvLLMollamaLangChainLlamaIndexRAGASUnslothAxolotlGradio

bonus depth

nanoGPT: train a GPT-2 from scratch on custom data (Karpathy)
Mixture of Experts (MoE): Mixtral architecture, routing
Long context: RoPE scaling (YaRN, LongRoPE), sliding window attention
Multimodal LLMs: LLaVA, Qwen-VL, image+text training

RL foundations

MDP: states, actions, rewards, transitions, discount factor γ
Value functions: V(s), Q(s,a), Bellman equations
Policy gradient theorem: REINFORCE, baseline variance reduction
Q-Learning, SARSA: TD methods, ε-greedy exploration
DQN: experience replay, target network, Atari breakthrough
Double DQN, Dueling DQN, Prioritized Experience Replay

modern deep RL

Actor-Critic: A2C, A3C — combine policy and value
PPO: clip objective, GAE, the algorithm that powers RLHF
SAC: entropy maximization, continuous action spaces
Model-based RL: Dyna, world models, Dreamer
Multi-agent RL: MARL basics, cooperative vs competitive

RL in LLMs

RLHF deep dive: reward model training, PPO on language models
GRPO: group relative policy optimization (DeepSeek-R1)
Constitutional AI: AI feedback instead of human feedback
Reward hacking: why reward shaping is hard

⬡ milestone

You can train a DQN to solve CartPole, implement PPO from scratch, explain how RLHF connects PPO to language model alignment, and train an agent on a custom Gym environment.

Project: RL Agent — Game or Custom Env

Train PPO with Stable-Baselines3 on a custom Gymnasium environment. Implement reward shaping. Compare DQN vs PPO. Visualize policies and value functions.

tools

gymnasiumstable-baselines3RLlibCleanRLMuJoCoIsaac Gymtrl (RL for LLMs)

bonus depth

AlphaGo / AlphaZero: MCTS + RL for game mastery
Offline RL: learning from logged data without online interaction
Sim-to-real transfer for robotics
Hierarchical RL: options framework, temporal abstraction

experiment tracking & data

Experiment tracking: MLflow, W&B — metrics, params, artifacts, comparisons
Data versioning: DVC — track datasets + models with Git-like semantics
Feature stores: Feast, Tecton — centralize feature computation
Data validation: Great Expectations, Evidently — catch schema drift
Dataset versioning: Hugging Face datasets, Delta Lake, LakeFS

model serving

ONNX: export PyTorch → universal format → optimize for hardware
TorchServe, Triton Inference Server: batching, GPU utilization
Quantization: INT8, FP16, dynamic vs static — inference speed vs accuracy
FastAPI model serving: async, Pydantic validation, batching
BentoML, Ray Serve: scalable model serving
Model registries: MLflow model registry, Hugging Face Hub

ml pipelines & monitoring

ML pipelines: Kubeflow Pipelines, ZenML, Metaflow
Orchestration: Airflow, Prefect for scheduled retraining
CI/CD for ML: test data pipelines, model quality gates in CI
Model monitoring: data drift (PSI, KS test), concept drift, prediction drift
Evidently AI, Arize, Fiddler for production monitoring
A/B testing, shadow deployment, champion-challenger
Model cards and Responsible AI documentation

⬡ milestone

You can take a research notebook, refactor it into a reproducible pipeline, track experiments with MLflow, serve the model as a FastAPI endpoint with ONNX, and monitor for data drift in production.

Project: End-to-End ML Pipeline

Full pipeline: data validation (Great Expectations) → training (MLflow) → evaluation → ONNX export → FastAPI serving → Evidently drift monitoring. Automated retraining trigger.

tools

MLflowW&BDVCZenMLBentoMLTritonONNX RuntimeEvidentlyKubeflowMetaflow

bonus depth

LLM-specific serving: vLLM, TGI (Text Generation Inference), Lorax
Edge deployment: CoreML (iOS), TFLite, TensorRT for NVIDIA GPUs
Continuous training: online learning, micro-batched updates

self-supervised learning

Contrastive learning: SimCLR, MoCo, BYOL — representations without labels
Masked autoencoders (MAE): BERT for images
CLIP-style: image-text contrastive at scale
Foundation models: pretraining → fine-tuning → probing
Data-efficient learning: few-shot, zero-shot, meta-learning
MAML: model-agnostic meta-learning

graph neural networks

Graph convolutions: GCN, GraphSAGE, GAT — message passing framework
Node classification, link prediction, graph classification
Knowledge graphs, recommendation systems with GNNs
PyG (PyTorch Geometric), DGL

multimodal AI

Vision-Language Models: LLaVA, Qwen-VL, Gemini architecture
Speech: Whisper (ASR), TTS systems, wav2vec 2.0
Video: VideoLLaMA, temporal understanding
Multimodal RAG: images + text in a single retrieval system

⬡ milestone

You can explain why self-supervised learning is so powerful, train a GNN on a graph dataset, fine-tune a vision-language model on a multimodal task, and read/implement recent papers from arXiv.

tools

PyTorch GeometricDGLopen_clipwhispertransformers (multimodal)

how to read papers

arXiv: cs.LG, cs.CL, cs.CV — follow daily; use Semantic Scholar
Papers With Code: track SOTA benchmarks and open implementations
Reproduce a paper's key result — the best way to deeply understand it

model evaluation rigorously

Benchmarking LLMs: MMLU, HumanEval, MATH, MT-Bench, LMSYS Arena
Evaluation pitfalls: benchmark contamination, metric gaming, p-hacking
Building custom eval suites for your domain
Human evaluation: annotation guidelines, inter-annotator agreement (Cohen's κ)
Confidence calibration: Expected Calibration Error, reliability diagrams

ai safety & alignment

Alignment problem: reward misspecification, Goodhart's law
Robustness: adversarial examples, distribution shift, OOD detection
Interpretability: attention visualization, probing classifiers, activation patching
Mechanistic interpretability: circuits, superposition (Anthropic research)
Red-teaming: systematic adversarial testing of LLMs

fairness & responsible AI

Bias in ML: measurement bias, label bias, historical bias, aggregation bias
Fairness metrics: demographic parity, equalized odds, calibration
Algorithmic auditing: fairness toolkits (Fairlearn, AIF360)
Privacy-preserving ML: federated learning, differential privacy
EU AI Act, model cards, datasheets for datasets

⬡ milestone

You can design a rigorous evaluation suite, audit a model for demographic bias, produce a model card following best practices, and explain the alignment problem to a non-technical stakeholder clearly.

tools

FairlearnAIF360SHAPLIMECaptum (PyTorch)LM-Eval-HarnessEleuther AI Evals

essential reading

Anthropic's Core Views on AI Safety — alignment research directly
"Concrete Problems in AI Safety" (Amodei et al.) — foundational paper
Distill.pub — interactive explanations of ML concepts

// specialization_paths

Choose Your Specialization Track

After the core modules, go deep on one track. These define your job title and differentiation.

LLM/GenAI Engineer

Build production LLM systems: RAG pipelines, fine-tuning, agents, multimodal apps. The hottest role in AI right now.

LLMsRAGAgentsvLLMLoRA

Computer Vision Engineer

Detection, segmentation, OCR, 3D vision, video understanding. High demand in autonomous vehicles, medical imaging, manufacturing.

YOLOSAMViTDiffusionONNX

NLP / Language Engineer

Semantic search, information extraction, document AI, text analytics. Strong in fintech, legal, and enterprise AI.

BERTNERFAISSspaCyHF

ML Engineer (Production)

Build, deploy, and operate ML systems at scale. MLOps, model serving, feature pipelines, retraining systems. High-paying and stable.

MLflowKubeflowTritonDVC

AI Research Scientist

Push the frontier — novel architectures, training methods, safety research. Requires deep math + ability to read and reproduce papers.

PapersJAXTheoryarXiv

Reinforcement Learning Engineer

Robotics, game AI, RLHF for LLMs, autonomous systems. Niche but extremely high-impact and less crowded than CV/NLP.

PPOSACMuJoCoRLHF

// curated_resources

The Best Resources — Named & Specific

Not generic suggestions. Each one is chosen for exactly what it teaches and where it fits in the roadmap.

Deep Learning (Goodfellow et al.)

The mathematical bible of deep learning. Covers MLP, CNN, RNN, optimization, regularization from first principles. Dense but authoritative.

Book · Free Online

Hands-On ML (Aurélien Géron)

The best practical textbook. Scikit-learn → TensorFlow/Keras. Every concept tied to code. Perfect companion for Modules 1–4.

Book · 3rd Edition 2023

The Elements of Statistical Learning

Hastie, Tibshirani, Friedman. Rigorous statistical theory behind ML algorithms. Free PDF. Read alongside Module 2 for depth.

Book · Free PDF

fast.ai (Jeremy Howard)

Top-down, practical-first deep learning. The best way to build intuition before diving into math. Do the first course before Module 3.

Course · Free

Andrej Karpathy (YouTube)

makemore, nanoGPT, micrograd — build language models and transformers from absolute scratch. The best neural net education on the internet.

YouTube · Free

Stanford CS231n (Vision) / CS224n (NLP)

World-class university courses. CS231n for CNNs and vision; CS224n for transformers and language. Lecture slides + assignments free online.

Course · Free

Hugging Face Course

The official course on transformers, tokenizers, fine-tuning, datasets. Hands-on, free, covers the entire HF ecosystem. Do after Module 5.

Course · Free

Papers With Code

Track SOTA results on any benchmark. Every paper linked to its open-source implementation. Essential for staying current from Module 6 onward.

Website · Free

Distill.pub

Interactive, beautifully illustrated deep dives on attention, feature visualization, interpretability. Highest quality ML writing that exists.

Website · Free

Kaggle Competitions

The real university of ML. Compete on tabular, vision, NLP challenges. Read winning solution writeups — gold mines of applied technique.

Platform · Free

The Batch (deeplearning.ai)

Andrew Ng's weekly ML newsletter. Clear explanations of the most important AI developments. Excellent for breadth and staying current.

Newsletter · Free

3Blue1Brown (YouTube)

The Neural Networks series and linear algebra/calculus playlists are the clearest mathematical intuitions you will find anywhere. Watch first.

YouTube · Free

// common_pitfalls

Mistakes That Derail AI/ML Learners

These are the real blockers — specific to AI/ML — that waste months of effort.

Skipping math to use frameworks directly

You can use PyTorch without understanding backprop. You cannot debug, improve, or innovate without it. Math is the moat.

Using test data during model development

Your test set is sacred. One peek and it's contaminated. Use a proper train/val/test split and only touch test at the very end.

Optimizing a model that hasn't been baselined

Always establish a simple baseline first (mean predictor, logistic regression). You can't know if your complex model is good without comparison.

Chasing leaderboard position instead of understanding

Rank 1 on MNIST teaches you nothing. Read winning Kaggle notebooks to understand techniques, then reproduce them on new problems.

Treating LLM outputs as ground truth

LLMs hallucinate confidently. Never use LLM outputs as training labels without human verification. Never trust an LLM's math without checking.

Ignoring data quality and jumping to model complexity

80% of real ML work is data. More data > better model, nearly always. Clean, well-labeled data beats a complex architecture on garbage data every time.

Not tracking experiments — running models and forgetting configs

Set up W&B or MLflow from your first experiment. Future you will desperately want to know what hyperparameters produced that good result 3 weeks ago.

Only learning PyTorch without understanding what it does

Implement one neural net in NumPy before touching any framework. This single exercise cements backprop better than any tutorial or book can.

Training on all data, evaluating on training data

A model that memorizes training data is useless. Generalization is the only thing that matters. Always evaluate on held-out data. Use k-fold on small datasets.

// progress_tracker

Module Completion Tracker

Check off each module as you complete the project milestone. Saved in your browser.

Completion Progress

0%(0 of 10 completed)

// capstone_project

Final Capstone Project

Choose a real problem in a domain you care about — medical imaging, code generation, document understanding, scientific data, or a Kaggle-style business problem. Build a complete AI pipeline: curate or collect data, perform rigorous EDA, establish baselines, train and evaluate multiple model families, apply the right fine-tuning technique (LoRA, full fine-tuning, or prompt engineering), optimize for inference (ONNX/quantization), serve via a REST API, add monitoring for drift, and document everything in a public GitHub repo + blog post or technical report. This is the project you show at interviews — it should demonstrate depth, not just familiarity.

Capstone - All 10 Modules

End-to-End AI System — Research to Production

Build something real, evaluate it rigorously, deploy it, and write about it publicly

Rigorous data pipeline

Baseline → SoTA comparison

Fine-tuned model (LoRA/full)

MLflow experiment tracking

ONNX optimized serving

Production monitoring

Model card + evaluation suite

Public blog post / paper

CoversAll 10 Modules

Duration4–8 weeks

DeliverableGitHub repo + live demo + writeup + model card

GoalSomething you're proud to show, not just completed

2025 Edition · Comprehensive Depth Roadmap

ArtificialIntelligence& Machine Learning

From linear algebra and Python to deep learning, transformers, MLOps, and building production AI systems — a rigorous, honest roadmap that doesn't skip the hard parts.

Start learning Track your progress

0Modules

0–18Months

0+Topics

0Projects

PythonLinear AlgebraStatisticsScikit-learnPyTorchTensorFlowTransformersLLMsComputer VisionNLPMLOpsReinforcement Learning

// prerequisites

Mathematical Foundations

This is what separates engineers who use ML from engineers who understand ML. Don't skip this. Study it in parallel with early modules.

Linear Algebra

Vectors, matrices, matrix multiplication, dot products, eigenvalues/eigenvectors, SVD, PCA. Everything in ML is linear algebra under the hood.

Statistics & Probability

Distributions (Gaussian, Bernoulli, Poisson), Bayes' theorem, MLE, MAP estimation, hypothesis testing, p-values, confidence intervals, CLT.

Calculus & Optimization

Derivatives, partial derivatives, chain rule, gradients, gradient descent, Hessians, Lagrange multipliers. Backprop is just the chain rule.

Information Theory

Entropy, cross-entropy, KL divergence, mutual information. These appear constantly in loss functions and model evaluation.

Numerical Methods

Floating point arithmetic, numerical stability, condition numbers. Understand why nan/inf appear and how to prevent them.

Graph Theory (Basics)

Nodes, edges, adjacency matrices, graph traversal. Essential for GNNs and understanding attention mechanisms as fully-connected graphs.

Key mathematical objects you'll use daily

Loss: L(θ) = -1/N · Σ [y·log(ŷ) + (1-y)·log(1-ŷ)]

Gradient: ∇θL = ∂L/∂θ

Update: θ ← θ - α · ∇θL

Attention: Attn(Q,K,V) = softmax(QKᵀ/√dₖ) · V

Bayes: P(θ|X) ∝ P(X|θ) · P(θ)

// learning_path

10-Module Deep Roadmap

Each module builds on the last. Core theory → implementation → project. Click any module to reveal hidden depth.

core topics

Python fluency: list comprehensions, generators, decorators, context managers
NumPy: ndarray, broadcasting, vectorized ops, einsum — the ML workhorse
Pandas: DataFrames, groupby, merge, apply, time-series indexing
Matplotlib + Seaborn: plotting distributions, correlations, model results
Jupyter notebooks: workflow, magic commands, widgets
SciPy: optimization, signal processing, stats
Virtual environments: venv, conda, poetry
Type hints, dataclasses, pydantic for clean ML code

⬡ milestone

You can load a CSV, clean messy data, run statistical summaries, visualize distributions and correlations, and write vectorized NumPy operations that process 1M rows in milliseconds — no loops.

Project: Exploratory Data Analysis Pipeline

Load a real dataset (Kaggle/UCI). Full cleaning, outlier detection, feature distributions, correlation heatmaps, and a written EDA report.

essential libraries

numpypandasmatplotlibseabornscipyplotlypolarspydantic

bonus depth

Polars vs Pandas: when to switch for 10x speed on large data
Dask: parallel computing for datasets that don't fit in RAM
Profiling Python code: cProfile, line_profiler, memory_profiler
C extensions and Numba for ultra-fast numerical code

supervised learning

Linear Regression: OLS, Ridge (L2), Lasso (L1), ElasticNet — derive the normal equation
Logistic Regression: sigmoid, cross-entropy loss, multi-class (softmax)
Decision Trees: ID3/CART algorithm, Gini impurity vs entropy, pruning
Random Forests: bagging, feature importance, OOB error
Gradient Boosting: XGBoost, LightGBM, CatBoost — the Kaggle weapons
SVM: hyperplane, kernel trick (RBF, polynomial), soft margin
k-NN: curse of dimensionality, distance metrics, approximate NN
Naive Bayes: generative model, MAP classification

unsupervised learning

K-Means, K-Medoids: initialization (K-Means++), convergence, elbow method
DBSCAN: density-based clustering, epsilon, min_samples
PCA: variance explained, scree plot, reconstruction
t-SNE, UMAP: high-dim visualization
Gaussian Mixture Models: EM algorithm

model evaluation (critical)

Bias-Variance tradeoff — the most important concept in ML
Cross-validation: k-fold, stratified, time-series CV
Classification metrics: precision, recall, F1, AUC-ROC, PR-AUC
Regression metrics: MAE, MSE, RMSE, R², MAPE
Confusion matrix, calibration curves, SHAP values
Hyperparameter tuning: GridSearch, RandomSearch, Optuna (Bayesian)
Feature engineering: encoding, binning, interactions, scaling
Imbalanced data: SMOTE, class weights, threshold tuning

⬡ milestone

Project: Tabular Prediction Competition

Solve a Kaggle tabular competition end-to-end: EDA, feature engineering, baseline → XGBoost → LightGBM ensemble. Beat baseline by >10%. Write a full model card.

tools

scikit-learnxgboostlightgbmcatboostoptunashapimbalanced-learnfeature-engine

bonus depth

Implement gradient descent for linear regression in pure NumPy
AutoML: AutoGluon, H2O.ai — understand what they do under the hood
Anomaly detection: Isolation Forest, LOF, Autoencoders
Survival analysis: Kaplan-Meier, Cox regression

from scratch first

Perceptron → MLP: forward pass, activation functions, output layer
Backpropagation: derive it fully from the chain rule — not optional
Activations: ReLU, Leaky ReLU, GELU, Swish, Sigmoid, Tanh — when to use each
Loss functions: MSE, BCE, CrossEntropy, Focal, Huber
Weight initialization: Xavier/Glorot, He, and why it matters
Vanishing/exploding gradients: causes and solutions

training techniques

Optimizers: SGD, Momentum, Adam, AdamW, Lion — intuition and math
Learning rate schedules: cosine annealing, warmup, OneCycleLR
Regularization: Dropout, L1/L2, BatchNorm, LayerNorm, WeightDecay
Batch size effects: large batch training, gradient accumulation
Early stopping, model checkpointing, training curves
Mixed precision training: fp16, bf16, torch.autocast

pytorch (go deep)

Tensors: creation, ops, broadcasting, GPU (.cuda() / .to(device))
Autograd: computational graph, .grad, retain_graph, no_grad
nn.Module: forward(), parameters(), buffers, hooks
DataLoader + Dataset: custom dataset, collate_fn, sampler
Training loop best practices: gradient clipping, AMP, DDP
torch.compile (PyTorch 2.x) for 2–3x speed
TensorBoard / W&B for experiment tracking

⬡ milestone

Project: Neural Net from Scratch + PyTorch Reproduction

Build a 3-layer MLP in NumPy (with backprop). Reproduce in PyTorch. Train on MNIST/CIFAR-10. Add BatchNorm, Dropout, LR scheduling. Compare all optimizers. Log to W&B.

tools

PyTorchTensorFlow/KerasWeights & BiasesTensorBoardLightningeinopstorchmetrics

bonus depth

Implement Adam optimizer from scratch in NumPy
Custom autograd: extend torch.autograd.Function for custom ops
JAX + Flax: functional approach to deep learning
torch.compile deep dive: Dynamo, TorchInductor

convolutional networks

Convolution operation: filters, stride, padding, dilation, groups
Pooling: max, average, global average pooling (GAP)
Architectures: LeNet → AlexNet → VGG → ResNet (skip connections!) → EfficientNet
Transfer learning: feature extraction vs fine-tuning vs linear probing
Data augmentation: RandomCrop, ColorJitter, Mixup, CutMix, AugMix
Image preprocessing: normalization, ImageNet stats, PIL vs OpenCV

object detection

R-CNN → Fast R-CNN → Faster R-CNN: region proposals, RoI pooling
YOLO (v5, v8): single-shot detection, anchor boxes, NMS
SSD, RetinaNet, DETR
mAP metric: IoU thresholds, PR curve per class
Semantic vs instance vs panoptic segmentation
SAM (Segment Anything): zero-shot segmentation

vision transformers

ViT: patch embeddings, positional encoding, class token
DeiT, Swin Transformer, ConvNeXt
CLIP: contrastive image-text pretraining — multimodal foundation
DINO, DINOv2: self-supervised visual features
Diffusion models for images: DDPM, stable diffusion architecture

⬡ milestone

You can fine-tune a ResNet50 on a custom dataset, train a YOLO model for object detection, achieve >90% accuracy on a 10-class problem, and implement a custom augmentation pipeline in Albumentations.

Project: Custom Object Detector

Collect/annotate data with Label Studio. Fine-tune YOLOv8. Evaluate with mAP. Export to ONNX. Deploy as a REST API with FastAPI.

tools

torchvisiontimmAlbumentationsultralytics (YOLO)OpenCVLabel StudioONNXDetectron2

bonus depth

Video understanding: 3D convolutions, optical flow, temporal models
Medical imaging: DICOM, 3D segmentation with nnU-Net
Generative models: GANs (StyleGAN), VAEs, Stable Diffusion fine-tuning (LoRA, DreamBooth)
Neural Radiance Fields (NeRF) for 3D scene representation

classical NLP

Text preprocessing: tokenization, lemmatization, stemming, stopwords
Bag of Words, TF-IDF, n-grams
Word2Vec, GloVe, FastText — static word embeddings
RNN, LSTM, GRU — sequential models and their problems
Seq2Seq with attention: the precursor to Transformers

transformers (deep)

Attention mechanism: queries, keys, values — derive from scratch
Multi-head attention, positional encoding (sinusoidal + RoPE)
Transformer architecture: encoder, decoder, encoder-decoder
BERT: masked LM, NSP, fine-tuning paradigm
GPT: autoregressive LM, causal attention mask
T5, BART: text-to-text and denoising
Tokenization deep dive: BPE, WordPiece, SentencePiece, tiktoken
Hugging Face: pipeline, AutoModel, Trainer API, datasets

NLP tasks

Text classification, sentiment analysis, NER, POS tagging
Question answering: extractive (SQuAD) and generative
Text generation: beam search, sampling, top-k, top-p, temperature
Summarization: extractive vs abstractive
Machine translation evaluation: BLEU, ROUGE, BERTScore
Semantic similarity: sentence-transformers, cosine similarity, FAISS

⬡ milestone

Project: Semantic Search Engine

Embed a document corpus with sentence-transformers. Build a FAISS vector index. Query interface with FastAPI. Fine-tune a classifier on domain-specific data. Evaluate with proper metrics.

tools

transformerstokenizerssentence-transformersFAISSspaCyNLTKdatasets (HF)evaluateaccelerate

bonus depth

Implement multi-head attention from scratch in PyTorch
Flash Attention 2: memory-efficient attention for long contexts
Efficient transformers: Longformer, BigBird, linear attention
Cross-lingual models: mBERT, XLM-R, multilingual embeddings

LLM architecture & training

Scaling laws: Chinchilla, compute-optimal training
Modern LLM architecture: RoPE, GQA, SwiGLU, RMSNorm, KV cache
Distributed training: data parallelism, tensor parallelism, pipeline parallelism (3D parallelism)
Training data: curation, deduplication, filtering (FineWeb, Dolma)
Instruction tuning: SFT on curated chat data
RLHF: reward model, PPO, Constitutional AI
DPO, ORPO, SimPO: preference optimization without RL

fine-tuning techniques

LoRA: low-rank adaptation — math intuition (W = W₀ + BA)
QLoRA: 4-bit quantization + LoRA for single-GPU fine-tuning
Prompt tuning, prefix tuning, adapter layers
Full fine-tuning vs PEFT trade-offs
Dataset formats: Alpaca, ShareGPT, ChatML
Evaluation: MMLU, HellaSwag, GSM8K, MT-Bench, custom evals

prompting & inference

Prompt engineering: zero-shot, few-shot, chain-of-thought, tree-of-thought
RAG: chunking, embedding, retrieval, reranking, generation
Quantization: GPTQ, AWQ, GGUF (llama.cpp)
Inference optimization: speculative decoding, continuous batching, PagedAttention (vLLM)
Tool calling / function calling: ReAct, structured outputs
Guardrails: output safety, factuality, hallucination detection
LLM agents: planning, memory, tool use (LangGraph, smolagents)

⬡ milestone

You can fine-tune Llama 3 on a custom task with QLoRA, evaluate it against a baseline, build a RAG system with reranking, and serve it with vLLM with streaming responses.

Project: Domain-Specific LLM + RAG System

Fine-tune Llama 3.1 8B on domain data with QLoRA. Build a RAG pipeline (LangChain + ChromaDB). Evaluate with RAGAS. Deploy with vLLM. Build a Gradio UI.

tools

transformerspefttrlbitsandbytesvLLMollamaLangChainLlamaIndexRAGASUnslothAxolotlGradio

bonus depth

nanoGPT: train a GPT-2 from scratch on custom data (Karpathy)
Mixture of Experts (MoE): Mixtral architecture, routing
Long context: RoPE scaling (YaRN, LongRoPE), sliding window attention
Multimodal LLMs: LLaVA, Qwen-VL, image+text training

RL foundations

MDP: states, actions, rewards, transitions, discount factor γ
Value functions: V(s), Q(s,a), Bellman equations
Policy gradient theorem: REINFORCE, baseline variance reduction
Q-Learning, SARSA: TD methods, ε-greedy exploration
DQN: experience replay, target network, Atari breakthrough
Double DQN, Dueling DQN, Prioritized Experience Replay

modern deep RL

Actor-Critic: A2C, A3C — combine policy and value
PPO: clip objective, GAE, the algorithm that powers RLHF
SAC: entropy maximization, continuous action spaces
Model-based RL: Dyna, world models, Dreamer
Multi-agent RL: MARL basics, cooperative vs competitive

RL in LLMs

RLHF deep dive: reward model training, PPO on language models
GRPO: group relative policy optimization (DeepSeek-R1)
Constitutional AI: AI feedback instead of human feedback
Reward hacking: why reward shaping is hard

⬡ milestone

You can train a DQN to solve CartPole, implement PPO from scratch, explain how RLHF connects PPO to language model alignment, and train an agent on a custom Gym environment.

Project: RL Agent — Game or Custom Env

Train PPO with Stable-Baselines3 on a custom Gymnasium environment. Implement reward shaping. Compare DQN vs PPO. Visualize policies and value functions.

tools

gymnasiumstable-baselines3RLlibCleanRLMuJoCoIsaac Gymtrl (RL for LLMs)

bonus depth

AlphaGo / AlphaZero: MCTS + RL for game mastery
Offline RL: learning from logged data without online interaction
Sim-to-real transfer for robotics
Hierarchical RL: options framework, temporal abstraction

experiment tracking & data

Experiment tracking: MLflow, W&B — metrics, params, artifacts, comparisons
Data versioning: DVC — track datasets + models with Git-like semantics
Feature stores: Feast, Tecton — centralize feature computation
Data validation: Great Expectations, Evidently — catch schema drift
Dataset versioning: Hugging Face datasets, Delta Lake, LakeFS

model serving

ONNX: export PyTorch → universal format → optimize for hardware
TorchServe, Triton Inference Server: batching, GPU utilization
Quantization: INT8, FP16, dynamic vs static — inference speed vs accuracy
FastAPI model serving: async, Pydantic validation, batching
BentoML, Ray Serve: scalable model serving
Model registries: MLflow model registry, Hugging Face Hub

ml pipelines & monitoring

ML pipelines: Kubeflow Pipelines, ZenML, Metaflow
Orchestration: Airflow, Prefect for scheduled retraining
CI/CD for ML: test data pipelines, model quality gates in CI
Model monitoring: data drift (PSI, KS test), concept drift, prediction drift
Evidently AI, Arize, Fiddler for production monitoring
A/B testing, shadow deployment, champion-challenger
Model cards and Responsible AI documentation

⬡ milestone

You can take a research notebook, refactor it into a reproducible pipeline, track experiments with MLflow, serve the model as a FastAPI endpoint with ONNX, and monitor for data drift in production.

Project: End-to-End ML Pipeline

Full pipeline: data validation (Great Expectations) → training (MLflow) → evaluation → ONNX export → FastAPI serving → Evidently drift monitoring. Automated retraining trigger.

tools

MLflowW&BDVCZenMLBentoMLTritonONNX RuntimeEvidentlyKubeflowMetaflow

bonus depth

LLM-specific serving: vLLM, TGI (Text Generation Inference), Lorax
Edge deployment: CoreML (iOS), TFLite, TensorRT for NVIDIA GPUs
Continuous training: online learning, micro-batched updates

self-supervised learning

Contrastive learning: SimCLR, MoCo, BYOL — representations without labels
Masked autoencoders (MAE): BERT for images
CLIP-style: image-text contrastive at scale
Foundation models: pretraining → fine-tuning → probing
Data-efficient learning: few-shot, zero-shot, meta-learning
MAML: model-agnostic meta-learning

graph neural networks

Graph convolutions: GCN, GraphSAGE, GAT — message passing framework
Node classification, link prediction, graph classification
Knowledge graphs, recommendation systems with GNNs
PyG (PyTorch Geometric), DGL

multimodal AI

Vision-Language Models: LLaVA, Qwen-VL, Gemini architecture
Speech: Whisper (ASR), TTS systems, wav2vec 2.0
Video: VideoLLaMA, temporal understanding
Multimodal RAG: images + text in a single retrieval system

⬡ milestone

You can explain why self-supervised learning is so powerful, train a GNN on a graph dataset, fine-tune a vision-language model on a multimodal task, and read/implement recent papers from arXiv.

tools

PyTorch GeometricDGLopen_clipwhispertransformers (multimodal)

how to read papers

arXiv: cs.LG, cs.CL, cs.CV — follow daily; use Semantic Scholar
Papers With Code: track SOTA benchmarks and open implementations
Reproduce a paper's key result — the best way to deeply understand it

model evaluation rigorously

Benchmarking LLMs: MMLU, HumanEval, MATH, MT-Bench, LMSYS Arena
Evaluation pitfalls: benchmark contamination, metric gaming, p-hacking
Building custom eval suites for your domain
Human evaluation: annotation guidelines, inter-annotator agreement (Cohen's κ)
Confidence calibration: Expected Calibration Error, reliability diagrams

ai safety & alignment

Alignment problem: reward misspecification, Goodhart's law
Robustness: adversarial examples, distribution shift, OOD detection
Interpretability: attention visualization, probing classifiers, activation patching
Mechanistic interpretability: circuits, superposition (Anthropic research)
Red-teaming: systematic adversarial testing of LLMs

fairness & responsible AI

Bias in ML: measurement bias, label bias, historical bias, aggregation bias
Fairness metrics: demographic parity, equalized odds, calibration
Algorithmic auditing: fairness toolkits (Fairlearn, AIF360)
Privacy-preserving ML: federated learning, differential privacy
EU AI Act, model cards, datasheets for datasets

⬡ milestone

tools

FairlearnAIF360SHAPLIMECaptum (PyTorch)LM-Eval-HarnessEleuther AI Evals

essential reading

Anthropic's Core Views on AI Safety — alignment research directly
"Concrete Problems in AI Safety" (Amodei et al.) — foundational paper
Distill.pub — interactive explanations of ML concepts

// specialization_paths

Choose Your Specialization Track

After the core modules, go deep on one track. These define your job title and differentiation.

LLM/GenAI Engineer

Build production LLM systems: RAG pipelines, fine-tuning, agents, multimodal apps. The hottest role in AI right now.

LLMsRAGAgentsvLLMLoRA

Computer Vision Engineer

Detection, segmentation, OCR, 3D vision, video understanding. High demand in autonomous vehicles, medical imaging, manufacturing.

YOLOSAMViTDiffusionONNX

NLP / Language Engineer

Semantic search, information extraction, document AI, text analytics. Strong in fintech, legal, and enterprise AI.

BERTNERFAISSspaCyHF

ML Engineer (Production)

Build, deploy, and operate ML systems at scale. MLOps, model serving, feature pipelines, retraining systems. High-paying and stable.

MLflowKubeflowTritonDVC

AI Research Scientist

Push the frontier — novel architectures, training methods, safety research. Requires deep math + ability to read and reproduce papers.

PapersJAXTheoryarXiv

Reinforcement Learning Engineer

Robotics, game AI, RLHF for LLMs, autonomous systems. Niche but extremely high-impact and less crowded than CV/NLP.

PPOSACMuJoCoRLHF

// curated_resources

The Best Resources — Named & Specific

Not generic suggestions. Each one is chosen for exactly what it teaches and where it fits in the roadmap.

Deep Learning (Goodfellow et al.)

The mathematical bible of deep learning. Covers MLP, CNN, RNN, optimization, regularization from first principles. Dense but authoritative.

Book · Free Online

Hands-On ML (Aurélien Géron)

The best practical textbook. Scikit-learn → TensorFlow/Keras. Every concept tied to code. Perfect companion for Modules 1–4.

Book · 3rd Edition 2023

The Elements of Statistical Learning

Hastie, Tibshirani, Friedman. Rigorous statistical theory behind ML algorithms. Free PDF. Read alongside Module 2 for depth.

Book · Free PDF

fast.ai (Jeremy Howard)

Top-down, practical-first deep learning. The best way to build intuition before diving into math. Do the first course before Module 3.

Course · Free

Andrej Karpathy (YouTube)

makemore, nanoGPT, micrograd — build language models and transformers from absolute scratch. The best neural net education on the internet.

YouTube · Free

Stanford CS231n (Vision) / CS224n (NLP)

World-class university courses. CS231n for CNNs and vision; CS224n for transformers and language. Lecture slides + assignments free online.

Course · Free

Hugging Face Course

The official course on transformers, tokenizers, fine-tuning, datasets. Hands-on, free, covers the entire HF ecosystem. Do after Module 5.

Course · Free

Papers With Code

Track SOTA results on any benchmark. Every paper linked to its open-source implementation. Essential for staying current from Module 6 onward.

Website · Free

Distill.pub

Interactive, beautifully illustrated deep dives on attention, feature visualization, interpretability. Highest quality ML writing that exists.

Website · Free

Kaggle Competitions

The real university of ML. Compete on tabular, vision, NLP challenges. Read winning solution writeups — gold mines of applied technique.

Platform · Free

The Batch (deeplearning.ai)

Andrew Ng's weekly ML newsletter. Clear explanations of the most important AI developments. Excellent for breadth and staying current.

Newsletter · Free

3Blue1Brown (YouTube)

The Neural Networks series and linear algebra/calculus playlists are the clearest mathematical intuitions you will find anywhere. Watch first.

YouTube · Free

// common_pitfalls

Mistakes That Derail AI/ML Learners

These are the real blockers — specific to AI/ML — that waste months of effort.

Skipping math to use frameworks directly

You can use PyTorch without understanding backprop. You cannot debug, improve, or innovate without it. Math is the moat.

Using test data during model development

Your test set is sacred. One peek and it's contaminated. Use a proper train/val/test split and only touch test at the very end.

Optimizing a model that hasn't been baselined

Always establish a simple baseline first (mean predictor, logistic regression). You can't know if your complex model is good without comparison.

Chasing leaderboard position instead of understanding

Rank 1 on MNIST teaches you nothing. Read winning Kaggle notebooks to understand techniques, then reproduce them on new problems.

Treating LLM outputs as ground truth

LLMs hallucinate confidently. Never use LLM outputs as training labels without human verification. Never trust an LLM's math without checking.

Ignoring data quality and jumping to model complexity

80% of real ML work is data. More data > better model, nearly always. Clean, well-labeled data beats a complex architecture on garbage data every time.

Not tracking experiments — running models and forgetting configs

Set up W&B or MLflow from your first experiment. Future you will desperately want to know what hyperparameters produced that good result 3 weeks ago.

Only learning PyTorch without understanding what it does

Implement one neural net in NumPy before touching any framework. This single exercise cements backprop better than any tutorial or book can.

Training on all data, evaluating on training data

A model that memorizes training data is useless. Generalization is the only thing that matters. Always evaluate on held-out data. Use k-fold on small datasets.

// progress_tracker

Module Completion Tracker

Check off each module as you complete the project milestone. Saved in your browser.

Completion Progress

0%(0 of 10 completed)

// capstone_project

Final Capstone Project

Capstone - All 10 Modules

End-to-End AI System — Research to Production

Build something real, evaluate it rigorously, deploy it, and write about it publicly

Rigorous data pipeline

Baseline → SoTA comparison

Fine-tuned model (LoRA/full)

MLflow experiment tracking

ONNX optimized serving

Production monitoring

Model card + evaluation suite

Public blog post / paper

CoversAll 10 Modules

Duration4–8 weeks

DeliverableGitHub repo + live demo + writeup + model card

GoalSomething you're proud to show, not just completed