SmartPrompt DRONA AI Tutor
← Hub 🎓 Academy 📋 BA Tools
Complete path

The DRONA Curriculum

12 phases. 45+ modules. Everything from "what is Python" to "deploy a 10-agent production system". Click any phase to see the full breakdown.

12
Phases
45+
Modules
200+
Topics
40+
Technologies
12–18
Months (10hr/wk)
Realistic time expectations

At 10 hours/week: 6 months to first working agent, 12–18 months to expert level. But you don't need to finish everything before you're useful. Most developers become productive enough to build real things after just Phase 5. DRONA adapts your path to your goals — you decide how deep to go.

All 12 phases

P00
Phase 00: Prerequisites
2–4 weeks
3 modules

The foundation. Before learning AI, you need two things: Python programming basics and some mathematics. Don't skip this — it's what makes everything else make sense.

📦 Python for AI

Learn Python specifically for AI work: variables, loops, functions, classes, list comprehensions, generators, file handling, and API calls.

PythonJupyter NotebookVSCodenumpypandas
📦 Mathematics for ML

The math that AI is built on: vectors and matrices (linear algebra), rates of change (calculus), and uncertainty (probability). Each explained from scratch with real examples.

Linear algebraCalculusProbabilityStatisticsnumpymatplotlib
📦 Data Manipulation

Working with data before feeding it to a model: cleaning, transforming, visualizing, and understanding what your data is telling you.

pandasnumpymatplotlibseaborndata cleaning
P01
Phase 01: Machine Learning Foundations
4–6 weeks
4 modules

The core ideas that all of modern AI is built on. How machines learn from examples. Why training works. What can go wrong and how to fix it.

📦 What is Machine Learning

Understand the different types of learning (supervised, unsupervised, reinforcement), the training loop, and the fundamental problem of overfitting.

scikit-learnmatplotlibnumpy
📦 Supervised Learning Algorithms

The classic algorithms: linear regression, logistic regression, decision trees, random forests, XGBoost. Build each from scratch, then use the library version.

Linear regressionLogistic regressionDecision treesRandom forestsXGBoostscikit-learn
📦 Unsupervised Learning

Find patterns without labels: K-Means clustering, PCA (dimensionality reduction), t-SNE (visualization). Essential for real-world data.

K-MeansPCAt-SNEUMAPscikit-learn
📦 Model Evaluation

The science of testing whether your model is actually good: cross-validation, precision/recall, ROC curves, hyperparameter tuning.

Cross-validationConfusion matrixAUC-ROCOptunascikit-learn
P02
Phase 02: Deep Learning
6–8 weeks
5 modules

Neural networks. The technology behind everything from image recognition to ChatGPT. Build one from scratch in numpy, then learn PyTorch, then understand CNNs and RNNs.

📦 Neural Networks from Scratch

Build a complete neural network using only numpy. Understand every line: forward pass, loss calculation, backpropagation, gradient descent.

numpyMathematicsBackpropagationGradient descent
📦 PyTorch Deep Dive

Master PyTorch: tensors, automatic differentiation, nn.Module, custom layers, GPU acceleration, training loops, model saving.

PyTorchCUDAautogradDataLoader
📦 CNNs — Convolutional Neural Networks

How computers see images. Convolution, pooling, the famous architectures (ResNet, VGG), transfer learning.

PyTorchtorchvisionResNetTransfer learningtimm
📦 RNNs — Recurrent Neural Networks

How AI processes sequences (text, time series, audio). LSTM, GRU, the vanishing gradient problem and how to solve it.

PyTorchLSTMGRUSequence modeling
📦 Regularization & Optimization

Keep your model from memorizing instead of learning: dropout, batch normalization, learning rate scheduling, mixed precision.

PyTorchBatch normDropoutAdam optimizer
P03
Phase 03: Transformer Architecture
4–6 weeks
5 modules

The breakthrough that created ChatGPT, GPT-4, and everything like them. Understand attention — the single most important idea in modern AI.

📦 Attention Mechanisms

The core idea: instead of reading text left-to-right, allow every word to 'attend' to every other word. Build scaled dot-product attention from scratch.

AttentionQueries-Keys-ValuesScaled dot-productPyTorch
📦 The Transformer Architecture

The full architecture of the paper 'Attention is All You Need' (2017). Encoder, decoder, positional encoding. Build a complete transformer.

TransformersEncoder-DecoderPositional encodingPyTorch
📦 BERT — Encoder Models

BERT and its family: how to use pre-trained language models for classification, named entity recognition, question answering.

BERTRoBERTaDistilBERTHugging FaceFine-tuning
📦 GPT — Decoder Models

How GPT and ChatGPT work: autoregressive language modeling, sampling strategies (temperature, top-p), few-shot learning.

GPTAutoregressiveSamplingOpenAI APIHugging Face
📦 Modern LLM Architectures

LLaMA, Mistral, Phi, Gemma: what makes them different. Grouped query attention, RoPE positional embeddings, Flash Attention, KV cache.

LLaMAMistralRoPEFlash AttentionKV cachevLLM
P04
Phase 04: LLM Training & Adaptation
4–6 weeks
3 modules

How to make an LLM do what you want. Fine-tuning, RLHF (how ChatGPT was created), and the revolutionary LoRA technique that makes it affordable.

📦 Fine-tuning Techniques

Full fine-tuning vs LoRA vs QLoRA. When to fine-tune vs prompt engineer. The trade-offs between cost, quality, and speed.

LoRAQLoRAPEFTHugging Face TRLUnsloth
📦 RLHF & Alignment

How ChatGPT was created from GPT-3.5. Reward modeling, PPO optimization, DPO (the simpler alternative). The alignment problem.

RLHFPPODPOReward modelingConstitutional AITRL
📦 Prompt Engineering (Complete)

Every prompting technique: zero-shot, few-shot, chain of thought, tree of thoughts, ReAct, self-consistency, JSON mode.

CoTFew-shotTree of ThoughtsReActDSPyLangChain
P05
Phase 05: Agentic AI Fundamentals
4–6 weeks
4 modules

The shift from AI that answers questions to AI that does tasks. The agent loop, tool use, memory, and planning — the four pillars of agentic AI.

📦 What is an AI Agent

The difference between a chatbot and an agent. The OODA loop. The four pillars: perception, reasoning, action, memory. When agents are worth using.

Agent architectureReActOODA loopTool use
📦 Tool Use & Function Calling

How agents interact with the world: OpenAI function calling, Anthropic tool use, JSON schema design, parallel tool calling, error handling.

OpenAI APIAnthropic APIFunction callingJSON Schema
📦 Agent Memory

The four types of memory in agents: working (context window), episodic (vector stores), semantic (LLM knowledge), procedural (tools).

Vector databasesLangChain memoryEntity memorySummarization
📦 Planning & Task Decomposition

How agents break down complex tasks: chain-of-thought planning, Plan-and-Execute, Tree of Thoughts, recursive decomposition.

Plan-and-ExecuteTree of ThoughtsTask decompositionLangGraph
P06
Phase 06: Agent Frameworks
4–6 weeks
4 modules

The tools that make building agents faster. Deep dives into LangChain, LangGraph, AutoGen, and CrewAI — when to use each and why.

📦 LangChain Complete

The most popular agent framework: chains, agents, tools, memory, LCEL. Build a full RAG pipeline with a tool-using agent.

LangChainLCELLangSmithRAGAgentsTools
📦 LangGraph — Stateful Workflows

For complex, production-grade agents: state machines, conditional branching, checkpointing, human-in-the-loop, streaming.

LangGraphState machinesCheckpointingStreamingHITL
📦 AutoGen — Multi-Agent Conversations

Microsoft's framework: agents that talk to each other. AssistantAgent, UserProxy, GroupChat, code execution sandboxes.

AutoGenAssistantAgentGroupChatCode execution
📦 CrewAI — Role-Based Teams

Define an AI team with roles (researcher, writer, analyst). Assign tasks. Watch them collaborate. Sequential and hierarchical workflows.

CrewAIAgent rolesTasksSequential processTools
P07
Phase 07: RAG & Knowledge Systems
4–6 weeks
4 modules

Retrieval Augmented Generation: how to make an LLM answer questions about YOUR documents, data, and knowledge bases. The most used agent pattern in production.

📦 Basic RAG

The core pipeline: load documents → chunk them → embed → store → retrieve → answer. Build it from scratch, then with LangChain.

LangChainLlamaIndexEmbedding modelsVector stores
📦 Vector Databases

The storage layer of RAG: Chroma (local), FAISS (fast), Pinecone (cloud), Weaviate (graph+vector), Qdrant (production).

ChromaFAISSPineconeWeaviateQdrantpgvector
📦 Advanced RAG Patterns

HyDE, multi-query retrieval, re-ranking, self-RAG, RAPTOR, GraphRAG. Go from 60% to 90% retrieval accuracy.

HyDERe-rankingSelf-RAGRAPTORGraphRAG
📦 Knowledge Graphs

Graph-based memory for agents: Neo4j, GraphRAG, entity extraction. When graphs beat vectors.

Neo4jGraphRAGCypherEntity extraction
P08
Phase 08: Multi-Agent Systems
4–6 weeks
3 modules

Multiple AI agents working together. How to design teams of agents, make them communicate, handle failures, and produce better results than any single agent.

📦 Architecture Patterns

Supervisor-worker, peer-to-peer, hierarchical, swarm. The trade-offs of each topology. When multi-agent is worth the extra complexity.

LangGraphAutoGenCrewAISupervisor pattern
📦 Communication Protocols

How agents talk to each other: message passing, shared state, A2A (Google's protocol), MCP (Anthropic's protocol).

A2A protocolMCPMessage passingShared state
📦 Emergent Behaviors

Debate agents, majority voting, speculative execution, self-play for improvement. Patterns that only work with multiple agents.

Debate patternMajority votingSpeculative executionAutoGen
P09
Phase 09: Orchestration & Workflows
4–6 weeks
4 modules

Making agents reliable in production: state machines, human oversight, error handling, retry logic, and professional workflow tools.

📦 Workflow Fundamentals

Sequential, parallel, conditional, event-driven workflows. DAG design. How to structure complex multi-step agent processes.

PrefectAirflowDAGsParallel execution
📦 Human-in-the-Loop (HITL)

When agents must pause and ask a human: approval gates, correction loops, confidence-based escalation. Implementation in LangGraph.

LangGraphInterruptHuman approvalEscalation
📦 Error Handling & Reliability

Every way agents fail and how to prevent it: retry, circuit breakers, fallback strategies, timeouts, idempotency.

Retry logicCircuit breakerFallbackTimeoutError recovery
📦 Production Workflow Tools

Prefect, Temporal.io, Celery for durable agent execution. When to use a dedicated orchestration tool.

PrefectTemporal.ioCeleryKafka
P10
Phase 10: Production & Advanced
6–8 weeks
5 modules

Taking agents from demo to production. Evaluation, monitoring, cost optimization, safety, and scaling. The difference between a toy and a real system.

📦 Agent Evaluation

How to know if your agent is actually good: AgentBench, GAIA, SWE-bench, LLM-as-judge, regression testing. Build your own eval framework.

LangSmithBraintrustAgentBenchGAIASWE-bench
📦 Observability & Monitoring

Trace every agent step, track cost per run, monitor latency and error rates. What to alert on and what to ignore.

LangSmithArize PhoenixW&BHeliconeTracing
📦 Safety & Security

Prompt injection attacks, privilege escalation, content filtering, guardrails. How to build agents that can't be manipulated.

Guardrails AINEMO GuardrailsPrompt injectionLLM Guard
📦 Scaling Agents

Stateless design, horizontal scaling, message queues for agent workers, Kubernetes for agent pools, cost optimization.

DockerKubernetesKafkaCeleryvLLMLiteLLM
📦 Real-World Case Studies

Build 5 complete production systems: customer support, code review, research assistant, document processor, data analyst.

Full-stackLangGraphFastAPIPostgreSQL
P11
Phase 11: Cutting Edge 2024–2025
Ongoing
5 modules

The frontier. What's happening right now in AI research and why it matters. These concepts will define the next 5 years.

📦 Advanced Agent Architectures

Mixture of Agents (MoA), Reflexion (agents that learn from mistakes), SELF-REFINE, Voyager (lifelong learning). The next wave.

MoAReflexionSELF-REFINEVoyagerResearch
📦 Code Generation Agents

SWE-agent, Devin-style architecture, sandboxed code execution, test-driven agents, GitHub integration.

SWE-agentE2BModalPlaywrightGitHub Actions
📦 Computer Use & GUI Agents

Claude Computer Use, browser automation, desktop agents. AI that can operate any software.

PlaywrightSeleniumClaude Computer UseBrowser agents
📦 Multimodal Agents

Vision + language agents, document understanding, audio agents, image generation as a tool.

GPT-4VClaude VisionDocument AIMultimodal
📦 The Future

A2A and MCP protocols, self-improving agents, the 50-year view. What to learn now to stay relevant.

A2AMCPFuture of AIResearch frontier

100 Real Examples Inside DRONA

DRONA contains 100 numbered examples — real, concrete illustrations of every key concept. Here are 12 examples from across the curriculum to show you the depth.

EX-016
Adam optimizer tracks both average gradient (direction) and average squared gradient (magnitude). Rare parameters, like rare words, get bigger updates. Common parameters get smaller ones.
EX-023
ResNet's skip connection: output = F(x) + x. If the block learns nothing useful, F(x)→0 and output = x (identity). This means useless layers become identity functions, solving the vanishing gradient problem in deep networks.
EX-027
Multi-head attention with 8 heads: each head learns different relationships. Head 1 might track subject-verb agreement; Head 2 tracks pronoun references. Nobody designed this — it emerges from training.
EX-040
Kaiming initialization: weights ~ Normal(0, √(2/n_in)). Designed for ReLU. Wrong initialization leads to either exploding or vanishing activations in the very first forward pass, before any training begins.
EX-057
LoRA: instead of updating 7B parameters, update two small matrices A and B where rank=16. Only 0.1% of parameters are trainable, but captures most of the fine-tuning benefit.
EX-063
Basic agent loop: while task not complete → get observation → LLM thinks → parse action → execute tool → append to history. The LLM is the reasoning engine, not the whole agent.
EX-071
The 20 most useful agent tools: web_search, code_execute, file_read/write, database_query, email_send, calendar_create, image_analyze, memory_store/recall, human_input, agent_spawn, validate_output.
EX-083
HyDE (Hypothetical Document Embeddings): instead of embedding the query, ask LLM to write a hypothetical answer, then embed THAT. Hypothetical answers are more similar to real answers in embedding space.
EX-090
Structured agent logging: every step records timestamp, agent_id, step_number, action_type, tokens used, latency, cost, success. Alert when cost >3× average or steps >2× average (stuck loop).
EX-094
A2A (Agent-to-Agent) Protocol (Google 2025): agents expose an Agent Card (capabilities, auth, pricing). One agent can discover, hire, and receive results from another. The beginning of an agent economy.
EX-098
Voyager (Wang et al. 2023): a lifelong learning agent in Minecraft. It discovered how to mine diamonds — 10 steps — through autonomous exploration. Nobody told it how. It found the solution itself.
EX-100
Neural scaling laws: Loss ∝ N^(-0.076) where N = parameters. This drove GPT-3 → LLaMA design: smaller models, massive data, efficient inference. Smaller models on more data beat larger models on less.

All 100 examples are inside DRONA

Copy DRONA, paste it into Claude, and type /examples [any topic] to see 5 examples for any concept.

Get DRONA Free → How to start