Project 01
LedgerLens — GST Invoice Intelligence API
Turns messy invoices and receipts — PDFs and phone photos — into validated, GST-ready data.
FastAPILangChainLangSmithPydantic
Vector 2.0 · Recorded course · Unlocks 21 Nov 2026
Become a Gen-AI Developer in 3 months
Even with Zero Gen-AI Experience.
Build and deploy five real AI products — RAG, agents, MCP, fine-tuning and LLMOps. All you need is Python, Docker & Git.
Fully recorded, with live support every Saturday
Content goes live 21 Nov 2026 · Enrol any time
- Videos
- 112
- Videos
- Real projects
- 5
- Real projects
- 1:1 mock interviews
- 5
- 1:1 mock interviews
- Every Saturday
- Live
- Every Saturday
Curriculum
What you'll learn
Eight sprints, from FastAPI to LLMOps. Tap a sprint to see what's inside.
You'll learn
- — HTTP and REST fundamentals — methods, status codes, headers, request and response bodies
- — FastAPI essentials — routing, path and query parameters, request bodies, response models and automatic API docs
- — Pydantic for validation — models, field validation, nested data and settings management
- — Async Python in practice — async/await, when async helps, and why one blocking call slows every user
- — Dependency injection — sharing database sessions, configuration and authentication cleanly across routes
- — Databases — Postgres with SQLAlchemy, schema design basics and migrations with Alembic
- — Authentication and security — API keys, JWT login, password hashing and protected routes
- — Background tasks and file uploads — safe uploads and work that runs after the response
- — Error handling, middleware and logging — consistent errors, request IDs and structured logs
- — Project structure and Docker — clean layout, env-based config, Postgres + Redis via docker compose
- — AI-assisted development with Claude Code, Cursor and GitHub Copilot — plan first, test always, review every change
Project shipped
Launchpad — Production FastAPI Starter Kit
A reusable, tested, Dockerised FastAPI template with authentication, database, migrations, logging and CI already wired in. Every project in the course starts from Launchpad.
You'll learn
- — How LLMs work — tokens, embeddings, attention, context windows and sampling, at the depth interviewers expect
- — LLM providers — OpenAI, Anthropic Claude and local open-source models with Ollama; when to choose which
- — The model landscape — generative, reasoning and decision models (System One models such as TypeSafe AI's Jev)
- — Prompt engineering for production — instructions, few-shot examples, delimiters and consistent behaviour
- — LangChain chat models, messages and content blocks with init_chat_model
- — Prompt templates — reusable, versionable prompts with variables and few-shot examples
- — Structured outputs — guaranteed Pydantic objects with with_structured_output, validation and retries
- — Tools and tool calling — @tool, binding tools to models and handling tool errors
- — Agents with create_agent — your first tool-using agent and the loop behind it
- — Middleware — PII masking, summarisation, call limits and model fallbacks without rewriting the agent
- — LCEL and runnables — composing with |, parallel steps, .with_retry() and .with_fallbacks()
- — Streaming, batching, conversation memory, caching and rate limiting
- — Vision inputs — extracting reliable data from images and scanned documents
- — Tracing with LangSmith and streaming responses through FastAPI with Server-Sent Events
- — Prompt injection basics and the first line of defence
Project shipped
LedgerLens — GST Invoice Intelligence API
An async FastAPI + LangChain service that reads invoices and receipts (PDFs and photos), extracts GST-ready structured data, validates it with business rules, and routes uncertain results to a human review queue.
You'll learn
- — RAG from first principles — embeddings, similarity search and why retrieval makes LLMs accurate and current
- — LangChain for RAG — loaders, text splitters, embedding models, vector stores and retrievers
- — Vector databases — Pinecone, pgvector and Qdrant compared; indexes, metadata filters, managed vs self-hosted
- — Ingestion pipelines — parsing PDFs, tables and scans with Docling; background workers; updates and deletions
- — Chunking strategies — why fixed-size fails, structure-aware chunking and retrieval-boosting metadata
- — Embedding model selection — quality, cost, dimensions and multilingual support for Indian languages
- — Hybrid search — BM25 plus vectors with reciprocal rank fusion for circular numbers, section IDs and product codes
- — Reranking with cross-encoders inside a strict latency budget
- — Query understanding — rewriting, multi-query expansion and conversational follow-ups
- — Grounded answers — citations, source attribution and saying "I don't know"
- — Multi-tenant and secure RAG — tenant isolation and document permissions enforced at query time
- — Evaluation — RAGAS and DeepEval metrics, trusted test datasets and CI quality gates
- — Scaling and cost — answer and embedding caching, latency budgets and scaling vector databases
Project shipped
RegRadar — Compliance Copilot for Indian Fintech
A multi-tenant RAG service that answers compliance questions from RBI and SEBI circulars plus each company's private policies, with precise citations, change tracking, and quality gates in CI.
You'll learn
- — Workflows vs agents — when a predictable workflow is right, and why most production systems are mostly workflow
- — LangGraph fundamentals — state, nodes, edges, conditional routing, cycles, parallel steps and subgraphs
- — State design — typed state, reducers, and what belongs in state vs your database
- — Durable execution — Postgres checkpointer so workflows resume exactly where they stopped after a crash
- — Human-in-the-loop — approve / edit / reject through your API and resume cleanly
- — Decision models in workflows — TypeSafe AI's Jev for classify, route and score steps with calibrated probabilities in milliseconds
- — Confidence thresholds so only uncertain cases reach a human or a bigger model
- — Queues and workers for high-volume processing with safe retries
- — Handling sensitive data — detecting and masking personal and health information with Presidio
Project shipped
ClaimSense — Health Insurance Claims Automation Engine
A durable, queue-driven LangGraph workflow that reads claim documents, checks them against policy rules, flags anomalies, calculates the payable amount, and brings in a human adjuster only when needed.
You'll learn
- — Agent fundamentals — the reason–act loop, tools, memory and goals
- — Agent design patterns — ReAct, planner–executor, reflection, supervisor and handoffs
- — Tool design — single-responsibility tools, clear schemas, idempotency and actionable errors
- — Loop control — step limits, cost ceilings, timeouts and graceful failure
- — Model Context Protocol (MCP) — tools, resources and prompts; stdio and Streamable HTTP; OAuth; MCP Inspector
- — MCP security — scoped permissions, tool poisoning and safe token handling
- — Agent2Agent (A2A) protocol — Agent Cards, tasks and streaming; MCP vs A2A
- — Multi-agent orchestration — supervisor and hierarchical patterns in LangGraph; shared vs isolated state
- — Multi-framework development — the same agent in CrewAI and Microsoft Agent Framework, and "why this framework?"
- — Guardrails and approvals — Guardrails AI or NeMo Guardrails, least-privilege tools, mandatory human approval
- — Red-teaming your own agent with promptfoo
- — A decision layer for agents — fast, bounded choices (which tool, retry or stop, escalate) in front of the LLM
- — Serving agents to many users — stateless servers, session storage, concurrency and streaming
Project shipped
TripPilot — Multi-Agent AI Travel Concierge
Tell it "Goa, 4 days, ₹40k, two people" — a team of AI agents searches flights, trains and hotels through MCP servers, checks your budget with a partner agent over A2A, builds a day-by-day itinerary, and books only after you approve.
You'll learn
- — When to fine-tune — prompting vs RAG vs fine-tuning, with cost and time trade-offs
- — How fine-tuning works — full vs parameter-efficient; LoRA and QLoRA
- — Preparing training data — cleaning, formatting, synthetic data and train/test splits
- — Hands-on fine-tuning with Hugging Face, PEFT, TRL and Unsloth on a real task from your projects
- — Evaluating and shipping your model — vs a larger model and RAG; merging adapters and pushing to the Hub
- — Three ways to make the same decision — prompted LLM vs fine-tuned model vs decision model, compared with data
- — AWS Bedrock — model access, IAM, the Converse API and using Bedrock from LangChain
- — Bedrock Knowledge Bases, Agents, Guardrails and model customisation
- — Bedrock cost and capacity — on-demand vs provisioned throughput, batch inference and prompt caching
- — Azure AI — Microsoft Foundry, model catalog and Azure OpenAI deployments from LangChain
- — Secure Azure access — managed identities, Key Vault, RBAC and private endpoints
- — Azure AI Search for RAG, Foundry Agent Service, Content Safety & Prompt Shields, Document Intelligence
- — Fine-tuning on Azure and cost & capacity — standard vs provisioned deployments, batch API and quotas
- — Choosing the right platform — Bedrock vs Azure AI vs direct APIs vs self-hosting, and which companies use what
Project shipped
Cloud AI Showdown
Fine-tune a small open model to classify ClaimSense documents, then run RegRadar's retrieval on both Bedrock Knowledge Bases and Azure AI Search, add Bedrock Guardrails and Azure AI Content Safety, and publish a side-by-side comparison of quality, speed and cost.
You'll learn
- — Deployment options — public APIs, Bedrock / Azure AI and self-hosted models, and cost per million tokens
- — Serving models with vLLM — your Sprint 5 fine-tuned model behind an OpenAI-compatible server; batching, quantization, benchmarks
- — Model gateway with the LiteLLM proxy — routing, fallbacks, team API keys and cost tracking
- — Planning for millions of users — 10M requests/day ≈ 116 RPS average and ~580 at peak
- — Load handling — per-customer rate limiting with Redis, exact / semantic / prompt caching, and queues
- — Decision models as a scale lever — moving classification, routing and verification off the LLM path
- — The circuit breaker pattern — closed, open and half-open states, thresholds, timeouts and shared state in Redis
- — Combining circuit breakers with retries, backoff and rate limits so they don't fight each other
- — Cloud deployment, your choice — AWS (Bedrock, ECS Fargate, ALB, ElastiCache, SQS, RDS, CloudWatch) or Azure (Azure OpenAI, AI Search, Container Apps, Service Bus, PostgreSQL, Monitor)
- — Kubernetes with autoscaling, CI/CD with GitHub Actions, evaluation gates, canary releases and fast rollback
Project shipped
TokenGrid — Million-User LLM Platform (part 1)
Serve your fine-tuned model, put a gateway with rate limiting, caching and circuit breakers in front of every model, deploy to AWS or Azure through CI/CD, and load-test it. TokenGrid is presented and defended in Sprint 7.
You'll learn
- — Capstone completion — finish TokenGrid and put RegRadar, ClaimSense or TripPilot behind it
- — Capstone architecture review — walked through exactly like a senior interview round
- — Present your platform — a demo and architecture walkthrough you can show any interviewer
- — Final mock interview — a full senior loop: resume deep-dive, system design, client case and behavioural
- — System design rounds — RAG at scale, an agentic workflow with human review, and an LLM platform for millions of users
- — Client-facing skills — discovery calls, scoping a pilot, success metrics and demos to non-technical stakeholders
- — Resume rewrite with honest, quantified outcomes, and keyword matching for each job description
- — GitHub portfolio polish — READMEs, architecture decision records, demo videos and pinned repos
- — LinkedIn optimisation, behavioural prep with STAR, and a salary negotiation playbook for India and remote roles
Project shipped
TokenGrid — completed, presented and defended
Your flagship platform, built across Sprints 6 and 7 — exactly how real engineering teams ship platforms — presented as a demo and defended in a senior-style architecture review.
What you ship
Five projects on your GitHub
Project 02
RegRadar — Compliance Copilot for Indian Fintech
Answers compliance questions from RBI and SEBI circulars and company policies — with exact citations.
FastAPILangChainPineconepgvector
Project 03
ClaimSense — Health Insurance Claims Automation Engine
A durable agentic workflow that processes thousands of claims a day without losing one.
LangGraphLangChainFastAPIPostgres checkpointer
Project 04
TripPilot — Multi-Agent AI Travel Concierge
Type "Goa, 4 days, ₹40k, two people" — a team of AI agents plans, prices and books the whole trip.
LangGraphMCP Python SDKA2ATypeSafe AI Jev
Project 05
TokenGrid — Million-User LLM Platform (Flagship)
The platform that serves, routes, protects and scales every LLM call — built for 10 million requests a day.
vLLMLiteLLMTypeSafe AI JevFastAPI
The stack
Tools that show up in real 2026 Gen-AI job descriptions.
All taught hands-on with production patterns — never as toy notebooks.
20 tools · 4 layers
FastAPI
Python API
LangChain
Agent framework
LangGraph
Stateful agents
Pydantic
Data validation
OpenAI
GPT models
Anthropic
Claude models
MCP
Tool protocol
A2A
Agent protocol
Jev
Decision model
HuggingFace
Open models
Pinecone
Vector DB
Qdrant
Vector DB
Postgres
+ pgvector
RAGAS
RAG evals
AWS Bedrock
Inference cloud
Azure AI
Enterprise AI
vLLM
Model serving
Kubernetes
Autoscaling
Docker
Containers
Redis
Cache + limits
+ production tooling you'll use
LangSmith
Langfuse
LiteLLM
CrewAI
Microsoft Agent Framework
Docling
DeepEval
Guardrails AI
promptfoo
Presidio
Ollama
OpenTelemetry
Prometheus
Grafana
k6 load tests
GitHub Actions
ECS Fargate
Azure Container Apps
Your instructor
Meet your instructor
Prateek Mishra
Sr. Gen-AI Developer Working at a global MNC
Founder & instructor · TechSimPlus
Verified instructor profile on LinkedIn
12000+
Engineers taught
8
Years Of Experience
20+
Production projects
I build production Gen-AI systems for a living and have shipped what most courses only describe in slides — agentic RAG pipelines on AWS Bedrock, MCP servers in production, multi-agent systems with stateful checkpointing.
Recorded, not alone
Learn at your pace. Get help when you need it.
Saturday Live Support
Stuck on something? Bring it to the live doubt session every Saturday.
5 one-on-one mock interviews
Real interview rounds with written feedback after each one.
Message the instructor
Send your questions directly — no waiting for the next class.
Your five mock interview rounds
- 01LLMs + Production RAGAfter Sprint 2
- 02Agentic Workflow DesignAfter Sprint 3
- 03Agents, MCP, A2A & SecurityAfter Sprint 4
- 04Cloud AI & ScaleAfter Sprint 6
- 05Final Senior LoopSprint 7
Pricing
One price. Everything included.
56% off · Vector 2.0
Vector 2.0
Complete recorded course · 112 videos
₹10,999
₹25,000
Save ₹14,001 · 56% offAll-inclusive course fee · pay in full or start with ₹5,999. No registration deadline — enrol any time.
Content goes live
21 Nov 2026
All sprints unlock
Live support
Every Saturday
Doubt sessions
Access
Lifetime
+ 5 mock interviews
Everything included
- 112 recorded videos — Pre-work + Sprints 0–7
- 5 production-grade Gen-AI projects
- AWS Bedrock, Azure AI and fine-tuning
- Saturday Live Support — weekly doubt sessions
- 5 one-on-one mock interviews with written feedback
- Direct messages to the instructor
- Vector City — a coding game for every video
- Resume, GitHub & LinkedIn guidance
- Private Discord community
- Lifetime access + certificate
3-day no-questions-asked refund · Secure checkout
112 videos. Eight sprints. One decision.
All content unlocks on 21 Nov 2026 · Enrol any time — no registration deadline.
About this course ⌄
About Vector 2.0 — the Gen-AI Developer course by Prateek Mishra
Vector 2.0 is a complete recorded, project-led online Gen-AI Developer course by TechSimPlus, taught by Prateek Mishra, a Sr. Gen-AI Developer with 8+ years of experience, 20+ production AI projects and 12,000+ engineers taught. It is the upgraded edition of Vector 1.0: 112 recorded videos across the pre-work and Sprints 0–7, with Saturday live support and Interview Support. All content unlocks on 21 Nov 2026. Registration has no end date — you can enrol any time. The course fee is ₹10,999 (down from ₹25,000, 56% off).
Across eight sprints you learn FastAPI, how LLMs work, LangChain, production RAG (Pinecone, pgvector, Qdrant, RAGAS, DeepEval), agentic workflows with LangGraph, AI agents with MCP and A2A, CrewAI and Microsoft Agent Framework, decision models, LLM fine-tuning with LoRA and QLoRA, AWS Bedrock, Azure AI and LLMOps — vLLM, LiteLLM, Kubernetes, caching, rate limiting and circuit breakers for millions of users. You ship five production projects — LedgerLens, RegRadar, ClaimSense, TripPilot and TokenGrid — plus the Cloud AI Showdown milestone and Vector City, a coding game with a graded stage for every video — and finish with system design and interview prep.
It is built for backend, software, data, QA and DevOps developers in India, the UAE, the UK, the US and Singapore who can write Python and want AI Engineer, Gen-AI Developer, Agentic AI Engineer, LLM Engineer or AI Platform Engineer roles. The only prerequisites are Python, Docker and Git. You learn at your own pace with lifetime access, and bring your doubts to the live support session every Saturday.