Lilly Tech Systems
AI Frameworks
Master the frameworks that power modern AI. 50 topics covering deep learning (PyTorch, TensorFlow, JAX, MLX), LLM/RAG (LangChain, LlamaIndex, DSPy, HuggingFace), distributed training (DeepSpeed, FSDP, Megatron, NeMo, Ray), classical ML (sklearn, XGBoost, LightGBM, pandas, Polars), specialized libraries (RAPIDS, PyG, spaCy, OpenCV, YOLO), and MLOps (MLflow, W&B, ClearML, DVC, Kubeflow, Ray).
The AI Frameworks track exists because the honest answer to "which framework should I use" has changed three times in the last five years, and because the cost of picking the wrong one (migration, retraining, team retraining, rewriting production code) is high. We cover PyTorch, TensorFlow, JAX, Keras, HuggingFace Transformers, LangChain, LlamaIndex, LangGraph, and the orchestration layers that sit above them, with attention to where each is strong and where it tends to disappoint teams.
What is useful about a framework comparison is rarely the feature list; it is the failure modes. Does the framework make it easy to profile and debug? Does it version well? Is the community active enough that a weird error has already been answered? Is the library small enough to reason about, or is it a stack of abstractions that hides the interesting parts? We apply that lens to every framework in the track so you can make the choice that will not haunt you in 18 months.
All Topics
50 topics organized into 6 categories spanning the full AI framework landscape.
Deep Learning Frameworks
PyTorch Mastery
Master PyTorch end-to-end. Learn tensors, autograd, nn.Module, DataLoader, torch.compile, distributed training, and the patterns that ship 80% of production AI today.
6 LessonsTensorFlow / Keras 3
Master TensorFlow 2.x and Keras 3 (multi-backend). Learn the functional API, tf.data, tf.function, distributed strategies, and TFLite for mobile/edge deployment.
6 LessonsJAX Mastery
Master JAX: XLA-compiled NumPy with autodiff. Learn jit, grad, vmap, pmap, sharding, and the patterns that power Google's most ambitious AI research.
6 LessonsFlax (JAX Neural Networks)
Master Flax: the most popular JAX neural network library. Learn nnx (new) and linen (old) APIs, training loops, and porting PyTorch models to Flax.
6 LessonsEquinox (JAX)
Master Equinox: PyTorch-style neural networks for JAX with full pytree compatibility. Learn modules, filtered jit, and the patterns for clean JAX research code.
6 LessonsPyTorch Lightning
Master PyTorch Lightning: organize PyTorch code into LightningModule, Trainer, and DataModule. Get distributed training, mixed precision, and checkpointing for free.
6 Lessonsfastai
Master fastai: high-level deep learning library on top of PyTorch. Learn DataBlock, Learner, callbacks, and the patterns that make state-of-the-art accessible.
6 LessonsApple MLX
Master Apple MLX: array framework optimized for Apple Silicon. Learn unified memory, lazy evaluation, MLX-LM, and the patterns for fast AI on M-series chips.
6 LessonsPaddlePaddle (Baidu)
Master PaddlePaddle: Baidu's open-source deep learning framework. Learn dynamic and static graphs, PaddleNLP, PaddleOCR, and the deployment story.
6 LessonsMindSpore (Huawei)
Master MindSpore: Huawei's open-source AI framework. Learn the AI native graph compiler, Ascend-optimized training, and when MindSpore beats alternatives.
6 LessonsLLM & RAG Frameworks
HuggingFace Transformers
Master HuggingFace Transformers: 1M+ pretrained models with one API. Learn AutoModel, AutoTokenizer, pipelines, Trainer, and the patterns for production HF use.
6 LessonsHuggingFace Diffusers
Master HuggingFace Diffusers: state-of-the-art diffusion models for image, video, and audio generation. Learn pipelines, schedulers, and custom training.
6 LessonsHuggingFace PEFT (LoRA, QLoRA)
Master HuggingFace PEFT: parameter-efficient fine-tuning. Learn LoRA, QLoRA, prefix tuning, prompt tuning, IA3, and the patterns for cheap LLM fine-tunes.
6 LessonsHuggingFace TRL (RLHF, DPO)
Master HuggingFace TRL: train LLMs with reinforcement learning. Learn SFTTrainer, DPOTrainer, PPOTrainer, KTOTrainer, and the alignment training patterns.
6 LessonsHuggingFace Accelerate
Master HuggingFace Accelerate: distributed training that 'just works'. Learn DeepSpeed/FSDP integration, mixed precision, and zero-code-change distributed launches.
6 LessonsLangChain
Master LangChain: the most popular LLM application framework. Learn LCEL, chains, retrievers, memory, and the patterns for production LangChain apps.
6 LessonsLlamaIndex
Master LlamaIndex: the data framework for LLM apps. Learn ingestion, indexing, query engines, agents, and the patterns for production RAG with LlamaIndex.
6 LessonsDSPy
Master DSPy: program LLMs declaratively, then optimize. Learn signatures, modules, optimizers (BootstrapFewShot, MIPRO), and the algorithmic prompt-engineering pattern.
6 LessonsHaystack
Master Haystack: production-ready LLM framework from deepset. Learn pipelines, components, document stores, and Haystack 2.0's modular architecture.
6 LessonsSemantic Kernel (Microsoft)
Master Microsoft Semantic Kernel: SDK for integrating LLMs into C#, Python, Java apps. Learn plugins, planners, and Microsoft's enterprise AI integration patterns.
6 LessonsTraining & Distributed
DeepSpeed
Master Microsoft DeepSpeed: train trillion-parameter models. Learn ZeRO stages, optimizer offload, pipeline parallelism, and the patterns for memory-efficient training.
6 LessonsPyTorch FSDP
Master PyTorch FSDP (Fully Sharded Data Parallel). Learn auto-wrap policies, mixed precision, activation checkpointing, and FSDP2 (per-parameter).
6 LessonsMegatron-LM (NVIDIA)
Master Megatron-LM: NVIDIA's framework for training huge language models. Learn tensor parallelism, sequence parallelism, and the patterns powering frontier-scale training.
6 LessonsNVIDIA NeMo
Master NVIDIA NeMo: end-to-end framework for conversational AI, ASR, TTS, and LLMs. Learn NeMo 2.0, NeMo Curator, NeMo Aligner, and production patterns.
6 LessonsRay Train
Master Ray Train: distributed training on Ray. Learn TorchTrainer, integration with FSDP/DeepSpeed, fault tolerance, and the patterns for elastic training.
6 LessonsComposer (MosaicML)
Master MosaicML Composer (now Databricks): efficient PyTorch training with algorithmic speedups. Learn Trainer, algorithms (e.g., ALiBi, EMA), and the deterministic recipes.
6 LessonsColossalAI
Master ColossalAI: efficient large-scale model training with auto-parallelism. Learn the Booster API, ZeRO, tensor parallelism, and ColossalChat for RLHF.
6 LessonsAxolotl
Master Axolotl: YAML-driven LLM fine-tuning framework. Learn config-driven training, LoRA/QLoRA recipes, and the patterns for fast iteration on 7B-70B models.
6 LessonsClassical ML
scikit-learn
Master scikit-learn: the bedrock of classical ML. Learn estimators, pipelines, cross-validation, hyperparameter search, and the patterns for production sklearn.
6 LessonsXGBoost
Master XGBoost: the gradient boosting framework that wins Kaggle. Learn DMatrix, parameters, GPU training, early stopping, and the production deployment patterns.
6 LessonsLightGBM
Master Microsoft LightGBM: faster gradient boosting via leaf-wise growth and histograms. Learn parameters, GPU support, and when LightGBM beats XGBoost.
6 LessonsCatBoost
Master Yandex CatBoost: gradient boosting with native categorical handling. Learn ordered boosting, target encoding, and when CatBoost shines on tabular data.
6 Lessonsstatsmodels
Master statsmodels: classical statistics in Python. Learn OLS, GLM, time series (ARIMA, SARIMAX), and the diagnostics and inference tools sklearn lacks.
6 LessonsNumPy
Master NumPy: the foundation of scientific Python. Learn ndarray, broadcasting, vectorization, einsum, and the patterns for fast numerical code.
6 Lessonspandas
Master pandas: dataframes for Python. Learn Series, DataFrame, groupby, merge, time series, and the patterns to handle 1M-1B rows efficiently.
6 LessonsPolars
Master Polars: blazing-fast Rust-based dataframes. Learn lazy vs eager, expressions, query optimization, and when Polars destroys pandas on speed.
6 LessonsSpecialized Frameworks
RAPIDS (cuDF, cuML)
Master NVIDIA RAPIDS: GPU-accelerated dataframes (cuDF) and ML (cuML). Learn pandas-like APIs at GPU speed and when RAPIDS gives 10-100x speedups.
6 LessonsPyTorch Geometric (PyG)
Master PyTorch Geometric: graph neural networks on PyTorch. Learn message passing, GCN, GAT, GraphSAGE, and the patterns for scalable graph learning.
6 LessonsDeep Graph Library (DGL)
Master DGL: framework-agnostic graph neural network library. Learn DGLGraph, message passing, sampling, and DGL on PyTorch/MXNet/TF.
6 LessonsspaCy (NLP)
Master spaCy: production-ready NLP in Python. Learn pipelines, transformers integration, custom components, and the patterns for fast NLP at scale.
6 LessonsNLTK
Master NLTK: the classic NLP library for Python. Learn tokenization, POS tagging, parsing, sentiment, and when NLTK still beats modern alternatives.
6 LessonsOpenCV
Master OpenCV: the dominant computer vision library. Learn image I/O, transformations, object detection, video, DNN module, and integration with PyTorch/ONNX.
6 LessonsDetectron2 (Meta CV)
Master Meta's Detectron2: state-of-the-art object detection and segmentation. Learn the model zoo, custom datasets, and production deployment patterns.
6 LessonsUltralytics YOLO
Master Ultralytics YOLO (v8, v11): blazing-fast object detection, segmentation, and pose estimation. Learn the unified API, training, and edge deployment patterns.
6 LessonsMLOps & Experimentation
MLflow
Master MLflow: the open-source MLOps platform. Learn tracking, projects, models, model registry, and the patterns to make ML reproducible at scale.
6 LessonsWeights & Biases (W&B)
Master Weights & Biases: experiment tracking, hyperparameter sweeps, artifacts, and reports. Learn how W&B becomes the system of record for ML teams.
6 LessonsClearML
Master ClearML: open-source MLOps platform. Learn experiment tracking, orchestration, datasets, model serving, and self-hosted MLOps patterns.
6 LessonsDVC (Data Version Control)
Master DVC: Git for data and ML pipelines. Learn dvc add, dvc push, pipelines (dvc.yaml), experiments, and the patterns for reproducible ML.
6 LessonsKubeflow
Master Kubeflow: ML toolkit for Kubernetes. Learn Pipelines, KServe (model serving), Katib (hyperparameter tuning), and the cloud-native MLOps stack.
6 LessonsRay (Distributed Compute)
Master Ray: distributed compute for AI. Learn Ray Core, Ray Data, Ray Tune, Ray Serve, Ray Train, and the patterns for scaling Python AI workloads.
6 LessonsWhy an AI Frameworks Track?
The right framework lets you focus on the problem; the wrong one becomes the problem.
Deep Learning Stack
PyTorch, TensorFlow, JAX, MLX, Flax, Equinox, Lightning, fastai, PaddlePaddle, MindSpore.
LLM Ecosystem
HuggingFace Transformers, Diffusers, PEFT, TRL, Accelerate; LangChain, LlamaIndex, DSPy, Haystack, Semantic Kernel.
Distributed Training
DeepSpeed, FSDP, Megatron-LM, NeMo, Ray Train, Composer, ColossalAI, Axolotl.
MLOps & More
Classical ML (sklearn, XGBoost, LightGBM, pandas, Polars), specialized (RAPIDS, PyG, spaCy, OpenCV, YOLO), MLOps (MLflow, W&B, ClearML, DVC, Kubeflow, Ray).