Gen AI Engineer – Agentic AI with Lang Graph, MCP & AWS Training in Hyderabad

1. Python Fundamentals 🐍

Goal: Master the core language mechanics every other module in this platform leans on. Gen AI Engineer

• 📦 Variables & Types
• 🛠️ Functions & Default Arguments
• 🏗️ Classes
• 🏷️ Dataclasses
• 📝 TypedDict
• 🔄 List & Dict Comprehensions
• 🔤 F-Strings
• ⚠️ Exceptions (try/except/finally)
• 🚪 The with Statement & Context Managers
• 💡 Type Hints
• 📁 pathlib
• 📜 Enums
• ✂️ String Methods
• 🔁 Iterators & Generators
• 🧩 JSON Serialization

2. Closures & Decorators 🎁

Goal: Understand how tool registration and cross-cutting behavior get wired without rewriting function bodies.

• 🔗 Closures & Free Variables
• 🔑 The nonlocal Keyword
• ⏳ Late Binding in Loops
• 🥇 First-Class Functions
• 🎀 Decorator Mechanics
• 🛡️ functools.wraps
• 🎛️ Decorators with Arguments
• 🥞 Stacking Decorators
• 🛠️ FastMCP @mcp.tool()/@mcp.resource()
• 🚀 FastAPI Route Decorators
• 🔀 Project Dispatch Closures
• 🔍 The @traced Decorator

3. Context Managers 🛡️

Goal: Guarantee cleanup — connections, sessions, checkpoints — even when something goes wrong.

• 🚪 with Statement Basics
• 🛑 __exit__ Exception Semantics
• 🛠️ contextlib.contextmanager
• ⚡ async with / __aenter__ / __aexit__
• 📚 AsyncExitStack
• 💾 SqliteSaver & LangGraph Checkpointing
• 🗄️ Manual sqlite3 Open/Close
• 🧪 FastAPI TestClient as a Context Manager
• 🪆 Nested & Multiple Context Managers
• 🐛 Spot-the-Bug: Context Manager Pitfalls

4. Regex 🔍

Goal: Detect PII, prompt injection, and intent with patterns that are precise and not catastrophically slow.

• 🔠 Character Classes & Quantifiers
• ⚓ Anchors & Word Boundaries
• 🎯 Groups: Capturing, Named & Non-Capturing
• 🔎 search / match / findall / finditer / sub / split
• ⚙️ Compiled Patterns & Flags
• 🐢 Greedy vs. Lazy Matching
• 👀 Lookahead & Lookbehind
• 🔀 Alternation & Backtracking
• 🛡️ Escaping & Raw Strings
• 🕵️ PII Detection Patterns
• 🚨 Prompt-Injection Detection Patterns
• 📧 Email Signature & Quoted-Reply Stripping
• 🛤️ Intent-Routing Regex Design
• 🔄 re.sub with Callback Functions
• 🔙 Backreferences
• 💥 Catastrophic Backtracking
• 🐛 Spot-the-Bug: Unescaped Characters & Missing Anchors
• 🐛 Spot-the-Bug: Off-by-One Quantifiers & Character-Class Ranges

5. Hashing, Tokens & Integrity 🔐

Goal: Generate secure tokens and detect tampering with cryptographic hashing.

• 🧮 SHA-256 Basics
• 🎲 secrets vs. random
• 🧩 Canonical JSON Hashing
• 🔗 Hash-Chained Audit Logs
• 🛡️ Tool Integrity & “Rug Pull” Detection
• 🎫 Scoped Credentials
• 🧼 Cryptographic Hygiene
• 🐛 Spot-the-Bug: Integrity Pitfalls
• 🔄 Project Hash Usage End to End

6. SQLite & FTS5 🗄️

Goal: Persist memory and search text at the storage layer — safely and efficiently.

• 🔌 sqlite3connect / cursor / execute
• 🛡️ Parameterized Queries & SQL Injection Prevention
• 🤝 Transactions & conn.commit()
• 🎣 fetchall() vs. fetchone()
• 🏗️ CREATE TABLE IF NOT EXISTS
• 🔑 PRIMARY KEY AUTOINCREMENT
• 📊 FTS5 Virtual Tables & UNINDEXED Columns
• 🔍 FTS5 MATCH Operator
• 🏆 bm25() Ranking Function
• ✂️ Porter Stemming Tokenizer
• 💻 Project Code: Memory / Calendar / Knowledge / SMTP Servers

7. Asyncio ⚡

Goal: Master the cooperative concurrency model every MCP client/server interaction is built on.

• ⏳ async def / await Fundamentals
• 🎡 The Event Loop
• 🚀 asyncio.run() & Top-Level Await
• 🧵 Coroutines vs. Tasks (create_task)
• 🤝 asyncio.gather Concurrency
• 🚦 asyncio.Event Signaling
• 🚪 Async Context Managers
• 📚 contextlib.AsyncExitStack
• ⏱️ Cancellation, Timeouts & Cancel Scopes
• 💻 Project Code: MCPClient._serve() / _run_loop

8. Threading & Sync/Async Bridge 🌉

Goal: Bridge a background asyncio event loop into a synchronous FastAPI request.

• 🧵 threading.Thread Basics
• 🚦 threading.Event
• 🔒 threading.Lock
• 🔄 Background Event Loop Pattern
• 🔀 run_coroutine_threadsafe
• 🔮 Predicting Cross-Thread Behavior
• 🐛 Spot-the-Bug: Thread/Async Pitfalls
• 📧 aiosmtpd Controller
• 🧬 MCPClient Lifecycle Management

9. FastAPI & Pydantic 🚀

Goal: Validate requests and serve the agent through a real, production-shaped web layer.

• 🏗️ BaseModel Field Definitions
• ⚙️ Optional Fields & Defaults
• 🚨 ValidationError
• 📤 response_model
• 🛣️ Path Parameters
• ❓ Query Parameters
• 📦 Body Parameters
• 🧪 TestClient
• 🚦 Status Codes
• 💉 Dependency Injection
• 📜 Enum Fields
• ✅ Field Validators
• 🪆 Nested Models
• 🐛 Silent Coercion Bugs
• 🛑 HTTPException
• 🧩 Model Serialization
• 🔌 End-to-End Request Integration

10. Building an MCP Server 🛠️

Goal: Expose tools and resources to a model through FastMCP’s decorator-driven API.

• 🏗️ FastMCP App Creation & @mcp.tool()
• 🧠 Schema Inference from Type Hints
• 📝 Docstrings as Tool Descriptions
• 📦 Return Types & structuredContent Wrapping
• 🔗 @mcp.resource() & URI Templates
• 📐 The Three MCP Primitives (Tool / Resource / Prompt)
• ⚖️ Choosing Tool vs. Resource
• 🔄 JSON-RPC Lifecycle: initialize → tools/list → tools/call
• ⚠️ isError vs. Protocol-Level Errors
• 🛡️ MCPToolError Handling
• 🧹 stdio Transport & Keeping stdout Clean
• 🔎 MCP Inspector
• ⚙️ Env-Var Configuration Patterns
• 🚫 HITL Gating Is Not Server-Side
• 🐛 Spot-the-Bug: Missing Type Hints & Mutable Defaults
• 🐛 Spot-the-Bug: Resource Decorated as a Tool
• 🐛 Spot-the-Bug: Swallowing Failure Without isError
• 🐛 Spot-the-Bug: Malformed URI Scheme
• 🏭 One Process Per Server
• 🔮 Predicting Schemas from Real Tool Signatures

11. Building an MCP Client 🤝

Goal: Drive the initialize → tools/list → tools/call handshake against a real subprocess.

• 🤝 ClientSession / stdio_client / StdioServerParameters
• 🔄 initialize → tools/list → tools/call Handshake
• 📚 AsyncExitStack for Multi-Session Clients
• 🌉 Sync Facade: Background Thread + run_coroutine_threadsafe
• 🚀 Subprocess Spawning: cwd / env / PYTHONPATH Pitfalls
• 🎁 structuredContent Unwrapping & Result Parsing
• ⚠️ MCPToolError / isError Handling
• 🐛 Real Bugs This Project Hit (Cancel Scope, ModuleNotFoundError)

12. LangGraph StateGraph 📊

Goal: Model an agent as a graph of nodes, edges, and durable state.

• 📊 StateGraph / TypedDict State Basics
• 🔗 add_node / add_edge
• 🔀 add_conditional_edges & Routing Functions
• 🎯 set_entry_point / compile() / invoke()
• 🛑 The END Sentinel
• 💾 Checkpointer / SqliteSaver Basics
• 💻 Project Code: gate / execute / rejected Nodes
• 🧵 get_state / thread_id / config

13. LangGraph Human-in-the-Loop ⏸️

Goal: Make an irreversible action genuinely, durably pause for human approval.

• ⏸️ interrupt() Basics
• 💾 SqliteSaver Durable Checkpointing
• ▶️ Command(resume=...) Resume Mechanics
• ⏭️ graph.get_state(config).next
• 🆔 thread_id Session Identity
• 🛡️ Cross-Process Durability Guarantee
• 🛑 Why HITL for Irreversible Actions
• 📝 HITLApprovalGraph.propose
• 🧩 HITLApprovalGraph.resume Edge Cases
• 📋 get_pending_action & list_pending_approvals
• 🚦 The Three States of an Action
• 🧠 Misconception: In-Memory Flag vs. SqliteSaver
• ❓ Misconception: Forgetting to Check state.next
• 🔄 Misconception: Resuming an Already-Completed Thread
• 🔮 Predicting Resume Behavior
• ✅ Correct Resume API Syntax
• 📐 HITL Design Rationale

14. BM25 & Hybrid Search 🔍

Goal: Rank documents by combining lexical scoring with semantic similarity.

• 📊 Term Frequency (Counter)
• 📉 Inverse Document Frequency (IDF)
• 🧮 The Full BM25 Formula
• 🗄️ SQLite FTS5 bm25()
• 📐 Cosine Similarity
• 🥇 Reciprocal Rank Fusion (RRF)
• 🏛️ Hybrid Search Architecture

15. Real Email & SMTP 📧

Goal: Send and receive genuinely real email through the SMTP protocol — not a simulation.

• ✉️ EmailMessage Construction
• 📤 smtplib.SMTP Usage
• 🗣️ The SMTP Protocol Conversation
• ⚙️ Env-Var Configuration
• 📥 aiosmtpd Controller & Local Persistence Testing

16. Observability & Tracing 🔍

Goal: See what your agents are actually doing in production, and gate regressions in CI.

• 📏 Span Dataclass
• 🌐 TraceContext & contextvars
• 🎀 The @traced Decorator
• 🚰 Dual-Mode Sink: In-Memory vs. LangFuse
• 📖 Reading Spans Back with get_sink()
• 🧪 EvalCase: One Row of an Eval Dataset
• 📊 MetricResult & EvalReport
• 🧮 Heuristic answer_correctness (F1 over Tokens)
• 🔀 evaluate_case’s prefer='auto' Routing
• 📚 run_eval_suite: Batching Evaluation
• 🚧 regression_gate: Comparing Against a Baseline
• ✅ gate_passed: The CI Build/Fail Decision

17. Token & Cost Optimization 💰

Goal: Spend fewer tokens for the same — or better — outcome, end to end.

• 🔑 Idempotency Keys & IdempotencyGuard.execute
• 💾 ResponseCache: Exact Match with TTL
• 🧠 ResponseCache: Semantic Match (Cosine)
• 📏 assert_prompt_budget
• 🧹 clean_pasted_text
• 📝 structured_fields_to_prompt
• 🛠️ project_tool_result
• 🛑 HardCaps & QuotaLedger.check_and_record
• 🚨 PerRequestMetrics.is_anomalous
• 🧾 CostAttributionLedger
• 💬 ConversationManager: keep_recent_turns & compact_at_tokens
• 🗜️ compress_context
• ✂️ extract_relevant_sentences
• 🚦 ModelRouter: Rules Pass Then Complexity Heuristic
• 📈 ModelRouter.log & simple_rate
• 🧩 structured_output_request & parse_structured_output
• ⚖️ MaxTokensPolicy & needs_retune
• 🏷️ metadata_prefilter
• 🎯 retrieve_wide_then_rerank
• 🌫️ vague_query_pipeline & recall_at_k

18. Chunking Strategies 🧱

Goal: Turn a long document into right-sized, retrievable pieces.

• 📦 Chunk Dataclass
• 🪟 fixed_size_chunk: Naive Word-Count Window + Overlap
• 📑 structure_aware_chunk: Splitting on Heading Boundaries
• 🪂 structure_aware_chunk: Fallback to Fixed-Size Within an Oversized Section
• 🗣️ semantic_sentence_window_chunk: Sentence-Boundary Chunking
• 📌 attach_heading: The Cheapest Relevance Boost in the File

19. Hallucination Defense 🛡️

Goal: Catch a model stating something its sources don’t actually support.

• 📋 GroundednessReport & is_grounded
• 🗂️ _content_words: Filtering Stopwords & Short Tokens
• 💯 groundedness_score: Per-Sentence Content-Word Overlap
• 🔎 verify_citations: Catching Fabricated Citation Indices
• ⚖️ evaluate_answer: Combining Groundedness & Citation Checks
• 🤐 should_abstain & guarded_answer

20. Guardrails & Prompt Injection Defense 🚧

Goal: Build the single entry/exit checkpoint every prompt and response passes through.

• ⚠️ InjectionRisk: Scoring Text into a Risk Level
• 📊 scan_for_injection: Weighted Pattern Categories
• 🤖 scan_for_injection: Blending in an Optional Model Classifier
• 🛑 InjectionScanResult.should_block
• 📦 wrap_untrusted_content: Structural Isolation
• 📜 build_instruction_hierarchy_preamble
• ⚖️ GuardrailAction, GuardrailDecision & is_blocked
• 🥇 PolicyRule & POLICY_RULES: First BLOCK Wins
• 🧪 evaluate_guardrails: Running the Pipeline End to End
• 🔄 run_input / output / tool_result_guardrails

21. RAG & Tool Poisoning Defense ☣️

Goal: Defend the two injection surfaces beyond the prompt: documents and tool definitions.

• 🎖️ TrustTier & _TIER_RANK
• 📄 DocumentProvenance & is_quarantined
• #️⃣ content_hash & build_provenance
• 🏥 quarantine_check: Reusing Injection Heuristics at Ingest Time
• 🔍 filter_by_trust: Narrowing What a Query Can Retrieve
• 🔒 verify_unchanged: Catching Post-Review Tampering
• 🛠️ ToolDefinition & definition_hash
• 🏗️ ToolIntegrityRegistry: Catching a “Rug Pull”
• 🧼 sanitize_description: Tool Descriptions Are Untrusted Content Too
• ✅ validate_call_arguments: Fail-Closed Schema Validation

22. Progressive Tool Discovery 🧭

Goal: Search over cheap tool summaries first, and never invoke what was never surfaced.

• 📝 ToolSummary: The Cheap, Always-Loaded Projection
• 🗂️ ToolRegistry.register: Building the Summary & Search Index
• 🔎 ToolRegistry.search: BM25 over Summaries (Stage 1)
• 💎 DiscoveredTool Dataclass
• 📈 _rerank: Stage 2, Scored Against the Full Description
• 🔦 DiscoverySession.search_tools: Recall Wide, Rerank Down
• 🚪 DiscoverySession.is_discovered & get: Closing the Direct-Call Loophole
• ✂️ DiscoverySession.context_snippet: Only K Tools Reach the Model

Platform Capstone Projects

Five projects lifted directly from genai_platform’s own modules. Each spans multiple curriculum modules and results in the same deployable,

production-shaped system already implemented in this repository — not a rebuild from scratch.

P1. Secure Multi-Server MCP Agent Gateway 🛡️

Scenario: An assistant that discovers and calls tools across four live MCP servers — memory, calendar, knowledge and email — without ever loading every tool schema into context.

• 🔌 FastMCP servers: memory_server, calendar_server, knowledge_server, email_server (stdio)
• 🧠 MCPBroker + MCPClient managing multiple concurrent stdio sessions
• 🔎 ToolRegistry + DiscoverySession (BM25 search → full-description rerank)
• 🎯 Two-stage discovery: cheap summary search first, full schema only for survivors
• 📚 AsyncExitStack holding every live session open across the request lifecycle
• 🌉 Sync facade (background thread + run_coroutine_threadsafe) for FastAPI callers
• 🚪 Direct-call loophole closed: a tool must be discovered before it can be invoked

P2. Human-in-the-Loop Approval Graph ⏸️

Scenario: An agent that proposes an irreversible action — sending an email, approving an expense — and pauses for a real human, resumable from a completely different process.

• 📊 LangGraph StateGraph (hitl/approval_graph.py) with gate / execute / rejected nodes
• 💾 SqliteSaver checkpointer for durable, cross-process state
• ⏸️ interrupt() and Command(resume=...) as the pause/resume primitives
• 🛑 Not a cosmetic “pending” flag — the graph genuinely cannot proceed unattended
• 🆔 thread_id-scoped state recoverable from any process that opens the same checkpoint DB
• 👀 get_state(config).next inspected to know exactly what is awaiting approval
• 🚦 Three explicit action states: pending, approved, rejected

P3. Enterprise RAG with Hallucination Defense 🏛️

Scenario: A knowledge-base Q&A tool that cites its sources, and refuses to answer with confidence it doesn’t actually have.

• 🧱 retrieval/chunking.py: structure-aware chunking with attach_heading
• 🔍 retrieval/hybrid_search.py: BM25 + cosine similarity + Reciprocal Rank Fusion
• 🛡️ retrieval/hallucination_guard.py: groundedness scoring & citation verification
• 📌 Heading-aware chunk attachment as a near-free relevance boost
• ⚖️ Hybrid lexical + semantic ranking instead of betting on either alone
• 💯 Per-sentence groundedness score against the actually-retrieved context
• 🤐 should_abstain() confidence gate returns an honest “I don’t know”

P4. Poisoning-Resistant Tool & Document Pipeline ☣️

Scenario: Locking down the two injection surfaces an attacker can reach besides the prompt itself: RAG documents and tool definitions.

• 🚧 security/prompt_injection.py + security/guardrails.py: scan, isolate, decide
• 📄 security/rag_poisoning.py: trust-tiered document provenance & quarantine
• 🛠️ security/tool_poisoning.py: tool-definition integrity registry
• 🔎 Weighted injection scanner + wrap_untrusted_content structural isolation
• 🥇 Ordered PolicyRule pipeline where the first BLOCK always wins
• 🎖️ DocumentProvenance trust tiers plus re-hashing to catch silent post-approval edits
• 🏗️ ToolIntegrityRegistry catches a tool quietly redefined after a human approved it

P5. Cost-Governed, Observable Production Agent 💰

Scenario: Running the whole stack in production with token budgets, per-team cost attribution, real tracing, and a CI gate that fails the build on a quality regression.

• ⚙️ optimization/: caching, cost_governance, model_router, history_and_compression
• 📡 observability/tracing.py: Span / TraceContext tracing built on contextvars
• 🧪 observability/eval_harness.py: EvalCase / EvalReport / regression_gate
• 💾 Idempotency-keyed response cache with exact and semantic (cosine) hits
• 🛑 HardCaps + QuotaLedger + CostAttributionLedger for per-team spend control
• 🚦 Cheap-first ModelRouter: rules pass, then a complexity heuristic
• ✅ gate_passed() turns a nightly eval run into an actual CI build/fail decision