ai

🤖 AI / LLM live

From Python backend to ML and AI engineer — 8 stages, fast track 16–20 weeks, full track 10–14 months

▸ Stage 0. Python for data (2–3 wk) live

  • NumPy: arrays, vectorization instead of loops, broadcasting
  • pandas or polars: loading, cleaning, groupby, merge, windows
  • Visualization and notebooks: matplotlib, Jupyter, uv environments
  • Formats and scale: Parquet, Arrow, DuckDB for SQL on files
  • Assignment: reproducible real-dataset EDA report

▸ Stage A. Math and statistics (6–8 wk, parallel) live

  • Linear algebra: vectors, matrices, norms, SVD idea, cosine similarity
  • Calculus: derivatives, gradients, chain rule as the base of backpropagation
  • Probability: Bayes, distributions, expectation, CLT
  • Statistics: MLE, confidence intervals, hypothesis testing, A/B tests
  • Optimization: gradient descent, SGD, momentum, Adam idea
  • Information theory: entropy, cross-entropy, KL-divergence
  • Assignment: linear regression from scratch on NumPy

▸ Stage B. Classical ML (8–10 wk) live

  • Tasks: regression, classification, clustering, dimensionality (PCA)
  • Models: linear/logistic, trees, gradient boosting (XGBoost, LightGBM, CatBoost)
  • Metrics: accuracy, precision, recall, F1, ROC-AUC vs PR-AUC on rare classes
  • Honest validation: CV, stratification, time splits, leakage detection
  • Practices: baselines first, sklearn Pipeline, Optuna, SHAP, reproducibility
  • Portfolio ML-1: tabular ML project with a leakage-free report

▸ Stage C. Deep learning (8–12 wk) live

  • PyTorch: tensors, autograd, nn.Module, training loop from scratch
  • Networks: MLP, CNN, RNN ideas; normalization, dropout, weight decay
  • Optimization: AdamW, LR schedules, gradient clipping, mixed precision
  • Transfer learning: pretrained weights, freeze layers, fine-tune
  • Transformers: tokenization, embeddings, self-attention, masks, encoder/decoder
  • Hugging Face ecosystem: Transformers, Datasets, PEFT, TRL
  • Experiment tracking: MLflow or Weights & Biases, seeds and configs
  • Portfolio ML-2: CNN experiments, text fine-tune, mini-GPT by hand

▸ Stage D. LLM and AI engineering (8–12 wk) live

  • LLM API craft: tokens, streaming, structured output, function calling, limits
  • Prompts as code; context engineering: what to keep in the window
  • Embeddings and vector stores: pgvector, Qdrant, FAISS, HNSW
  • RAG: chunking, hybrid search with RRF, reranking, citations, per-document ACL
  • Evals: 50–200 golden questions, LLM-as-judge, quality gates in CI
  • Agents: tool loop, step and budget limits, human confirmations, MCP
  • LLM security: prompt injection, OWASP Top 10 for LLM, least-privilege tools
  • Cost and latency: model routing, prompt caching, streaming, max_tokens
  • Fine-tuning when justified: LoRA/QLoRA, DPO; self-hosting with vLLM
  • Portfolio ML-3 (RAG with evals) and ML-4 (agent with injection defense)

▸ Stage E. MLOps and ML systems (6–10 wk) live

  • Reproducibility: code/data/model versioning (DVC), Docker, seeds
  • Experiment registry: MLflow or W&B; model lifecycle stages
  • Pipelines: Airflow, Prefect, Dagster; idempotent tasks and retries
  • Train-serving skew; shared feature code; feature stores (Feast)
  • Serving: FastAPI with ONNX/TorchScript, Triton, BentoML, vLLM
  • Rollouts: shadow, canary, A/B, automatic rollback
  • Monitoring: latency, data drift (PSI), alerts, prediction logging
  • Portfolio ML-5: deploy, drift monitoring, one-command retrain

▸ Stage F. Specializations (by choice) live

  • Recommenders: two-tower models, NDCG/MAP, cold start, feedback loops
  • Search and ranking: BM25, hybrid retrieval, learning-to-rank
  • LLM training: pretraining, SFT, RLHF/DPO, scaling laws, distributed runs
  • Computer vision: detection, segmentation, ViT, diffusion models
  • Time series: seasonality, ARIMA/Prophet vs boosting, time-aware validation
  • Antifraud: heavy imbalance, graph features, labeler feedback loops
  • Speech: ASR/TTS like Whisper, streaming latency
  • Inference infrastructure: GPU profiling, KV-cache, continuous batching, vLLM internals

▸ Stage G. GPU Inference in Kubernetes (vLLM, DRA/MIG), AI Gateways & MCP Ecosystem live

  • NVIDIA GPU Operator in Kubernetes: Dynamic Resource Allocation (DRA), MIG profiles, and Time-Slicing
  • High-throughput LLM serving with vLLM: PagedAttention, prefix caching, tensor parallelism, and 16Gi /dev/shm
  • Scale-to-zero GPU autoscaling with KEDA and rapid spot/on-demand GPU node provisioning via Karpenter
  • AI Gateways (LiteLLM and Envoy AI Gateway): token rate limits, team budgets, fallback routing, and Redis semantic cache
  • Model Context Protocol (awesome-mcp-servers) and IDE rules (.cursor/rules) for production coding agents

Source: roadmaps/ai.json · ETA: 2026-11