Skip to content
Launchpad
Enrol

T1· Engineering: tech

AI Engineering: LLM apps, agents & MCP

You will build, test and secure real LLM applications, agents and MCP servers on seller data, and prove they work with evals.

Choose a planRead the syllabus
Who it is for
CSE, AI/ML and ISE students who can write Python and want to build AI products, not just use them. Laptop required for the labs.
AI core
9 units · 20 h
This course
6 modules · 12 lessons · 2.3 h
Live projects
2, on Professional and Career
01Roles it prepares you for
  • AI Engineer (fresher)
  • GenAI Developer
  • LLM Application Developer

The skills these roles ask for. Preparation, not a promise of a job.

02Tools you will use
  • Claude API (Saarthi sandbox metered credits)Saarthi sandbox
  • Gemini API free tier (Google AI Studio key)Free tierVerified
  • OpenAI API (Saarthi sandbox metered credits)Saarthi sandbox
  • MCP Python SDK (mcp, MCPServer)Free tierVerified
  • MCP InspectorFree tierVerified
  • LangGraphFree tierVerified
  • CrewAIFree tier
  • PostgreSQL + pgvectorFree tierVerified
  • GitHub Copilot FreeFree tierVerified
  • Saarthi MCP server (anonymised seller data)Saarthi sandbox
  • Anonymised seller data pack v1 (Saarthi sandbox)Saarthi sandbox

You never pay for a tool subscription. “Verified” means we checked the free access on the vendor’s own site this month. Tool names describe what you learn; no vendor endorses Launchpad.

03Syllabus

What you will learn, in order.

Lesson films are narrated with an AI voice, over visuals made with AI tools and checked by our team. Saarthi, the coach, is an AI. Mentors and reviewers are people.

Part 1 · The AI core, shared by every stream

00Set up your coach1 hLearn how the core works, write the learning-context document that tells every AI tool who you are, and set Saarthi up as a coach that makes you think instead of thinking for you.
  1. How this core works6 min
  2. Write your learning-context document8 min
  3. Seven roles for AI, and study modes7 min
01How LLMs work and fail2.5 hBuild a working mental model of a large language model: tokens, prediction, context windows and memory. Then learn its predictable failures (hallucination, sycophancy, stale knowledge, bad arithmetic) and reproduce them yourself.
  1. Tokens and next-word prediction7 min
  2. Context windows and memory7 min
  3. Hallucination, sycophancy and other failures8 min
  4. Lab: one prompt, three models, five failures8 min
02Description and context engineering3.5 hWrite briefs that get the right output the first time: role, task, context, examples and format. Then make them reusable with Claude and ChatGPT Projects, Gemini Gems, custom GPTs and NotebookLM, and learn to work with PDFs, spreadsheets and images.
  1. The brief: role, task, context, format8 min
  2. Examples and output formats7 min
  3. Projects in Claude and ChatGPT8 min
  4. Gems, custom GPTs and NotebookLM8 min
  5. Working with files: PDFs, spreadsheets, images8 min
03Verify everything2.5 hCheck AI output the way fact-checkers do: lateral reading, tracing citations and numbers to the original, using deep-research modes without trusting them blindly, and testing a prompt properly with a 10-case golden set.
  1. Lateral reading7 min
  2. Checking citations and numbers8 min
  3. Deep-research modes: use and limits7 min
  4. Your first eval: a 10-case golden set8 min
04Connectors and MCP2 hConnect AI to the places your work lives (Drive, Gmail, Sheets, business data) without writing code. Learn what an MCP server is, why a reviewed directory differs from unreviewed community servers, how to grant the least access that works, and how to keep browser extensions on a short leash.
  1. What connectors and MCP are8 min
  2. Trust, least privilege and the reviewed directory8 min
  3. Browser extensions and AI browsers7 min
  4. Lab: ask the seller-data MCP server, check every answer8 min
05Agents and automations3.5 hHand multi-step work to AI safely. Tell an agent from an automation, decide what to delegate, use browser and coding agents with care, and build an n8n workflow that uses AI with a human-approval step before anything leaves the system.
  1. Agents vs automations: what to delegate8 min
  2. Browser agents, safely7 min
  3. Automation basics with n8n and Zapier8 min
  4. Lab: build an automation with a human-approval step8 min
  5. Failure modes and your failure-mode note7 min
06AI for learning and careers1.5 hUse AI to learn faster without outsourcing the learning, write applications that still sound like you, and practise interviews out loud with Saarthi.
  1. Learning with AI, not instead of learning7 min
  2. Applications in your own voice7 min
  3. Mock interview with Saarthi8 min
07Diligence: law, ethics, integrity2.5 hThe rules that apply when you use AI in India: the DPDP Act 2023 and its 2025 Rules, academic integrity and disclosure, copyright (including where Indian law is still unsettled), and bias. You finish by writing your own AI-use policy.
  1. DPDP Act 2023: personal data and AI8 min
  2. Academic integrity and disclosure7 min
  3. Copyright and AI in India8 min
  4. Bias and fairness7 min
  5. Write your personal AI-use policy7 min
08Core capstone1 hPut all four Ds together on one anonymised marketplace settlement file: ask good questions, automate one step, verify every number against Saarthi's P&L engine, and defend your work live in five minutes.
  1. The capstone brief6 min
  2. Lab: ask, automate, verify8 min
  3. Prepare your 5-minute defence7 min

Part 2 · AI Engineering: LLM apps, agents & MCP

01LLM APIs as a programming surface2 lessonsCalls, tokens, cost and structured output across Claude, OpenAI and Gemini, with the API as code rather than chat.
  1. What an API call really is9 min
  2. Lab: structured extraction on two providers12 min
02Tool use and agents2 lessonsThe agent loop, when a workflow beats an agent, and the budgets, permissions and approval gates every agent needs.
  1. How agents work, and when not to build one9 min
  2. Lab: a settlement agent that waits for approval14 min
03Building MCP servers2 lessonsMCP architecture, tool design for models, transports and trust, then a working server over the data pack.
  1. MCP from the server side9 min
  2. Lab: your first MCP server over seller data14 min
04Retrieval: embeddings, pgvector and RAG2 lessonsHow retrieval works and fails, when long context is simpler, and hybrid search with citations in Postgres.
  1. RAG that tells the truth9 min
  2. Lab: hybrid search with pgvector14 min
05Evals: golden sets and LLM-as-judge2 lessonsCode graders, calibrated LLM judges, trajectory checks for agents, and evals that run in CI.
  1. Evals you can trust9 min
  2. Lab: a 30-case eval with a calibrated judge15 min
06Shipping safely: security, cost and observability2 lessonsPrompt injection, data boundaries, spend caps and tracing, then a red-team suite against your own app.
  1. Prompt injection and the other ways LLM apps break9 min
  2. Lab: red-team your own copilot15 min
04Live projects

Two briefs. Real-shaped work.

  1. Project 1

    Seller-data MCP server with a 30-question eval

    Build an MCP server that exposes a seller's anonymised orders, settlements and returns to an AI host such as Claude. A seller should be able to ask plain-language questions like 'Which five SKUs lost the most money to returns in March?' and get correct, explained answers. Your server must pass a 30-question eval in which every numeric answer is checked against Saarthi's P&L engine output for the same data-pack version. The server must be safe to point at a real seller's data later: read-only, scoped to one seller, and resistant to instructions hidden in the data.

    You hand in

    • Git repository with the MCP server (Python or TypeScript SDK), at least five task-shaped tools and one resource
    • Streamable HTTP deployment on localhost or the sandbox, with token authentication that scopes every call to one seller
    • evals.jsonl with 30 cases (typical, edge, refusal and regression) and a one-command runner
    • Eval report: numeric pass rate as counts against engine outputs, trace checks, judge agreement with your own labels, and every failure explained
    • Three injection tests through data fields, with results
    • A 5-minute recorded demo and a live defence in which you answer questions on your design

    Built on Saarthi’s engines. Your key numbers are checked against Saarthi's own P&L logic, run on the same anonymised data.

    Data: Anonymised seller data pack v1: orders.csv, settlement_v2.csv, returns.csv, sku_master.csv, plus engine_outputs_v1.json (frozen Saarthi P&L engine outputs) for grading. Reviewed by Saarthi against a published rubric, signed off by a mentor.

  2. Project 2

    Marketplace policy copilot with citations and a red-team

    Build a RAG 'policy copilot' that answers sellers' questions about marketplace fees, returns and listing policies from a dated snapshot of public policy and fee pages. Every claim must cite a passage the reader can check, and the copilot must say so when the documents do not contain the answer. You will measure it with a golden-set eval and attack it with a prompt-injection red-team, including instructions planted inside the documents themselves.

    You hand in

    • Ingestion pipeline with heading-aware chunking, table handling and effective-date metadata
    • Hybrid retrieval (vector plus keyword) on pgvector or Chroma, with answers that cite chunk IDs checked in code
    • Golden set of at least 40 questions, including unanswerable and date-sensitive ones, with retrieval recall@5 and answer accuracy reported separately
    • Red-team suite of at least 20 attacks across five families, with before-and-after results
    • Cost and latency report per question, with an example monthly estimate clearly labelled as an estimate
    • Peer review of another student's copilot, plus a live defence

    Data: Dated snapshot of public marketplace fee and policy pages packaged in the sandbox, plus a synthetic fee card with planted conflicts for the red-team. Reviewed by Saarthi against a published rubric, signed off by a mentor.

05Practise out loud

In voice mock interviews and client calls, Saarthi plays: Priya, engineering lead at a Bengaluru e-commerce SaaS company, running a 45-minute technical round. She asks you to design a seller-support RAG system and an MCP server on a whiteboard, presses on how you would evaluate them and defend against prompt injection, and pushes back whenever you claim something works without saying how you measured it.

On plans that include mock interviews. You get written feedback after every call.

06Enrol

Pick a plan for AI Engineering: LLM apps, agents & MCP.

One payment. The access deadline for each plan is shown before you pay.

  • Foundation

    01

    Learn the AI your field now runs on.

    ₹2,999

    One payment. GST included.

    • Access for 120 days (about 4 months) from the day you pay
    • No 1:1 mentor sessions
    • The common AI core and your domain course
    • Saarthi, your AI coach, any hour
    • Hands-on labs on free tools
    • Verifiable certificate
    Enrol on Foundation

  • Professional

    02

    Prove it on real work.

    ₹7,999

    One payment. GST included.

    • Access for 180 days (about 6 months) from the day you pay
    • One 1:1 session with a mentor
    • Everything in Foundation
    • Two live projects, reviewed by Saarthi and a mentor
    • Public portfolio page of verified work
    • Voice mock interviews
    • One 1:1 coaching session
    Enrol on Professional

  • Career

    03

    Work alongside our team.

    ₹14,999

    One payment. GST included.

    • Access for 240 days (about 8 months) from the day you pay
    • Four 1:1 sessions with a mentor
    • One 1:1 call with an industry expert
    • Everything in Professional
    • Four-week supervised practicum with the Saarthi team
    • Four 1:1 coaching sessions with your mentor
    • An exclusive 1:1 call with an industry expert in your field
    • CV and LinkedIn review
    • Priority mentor replies
    Enrol on Career

Your access period starts the day you pay and is shown before you pay. One exam pause per enrolment, for up to 15 days. When access ends the course locks, and your portfolio page and certificate stay live. Full refund within 7 days, pro-rata after that, and an automatic refund for any practicum, 1:1 mentor or expert session we do not deliver. Refund policy