Launch
Capstone Project 3
Capstone Project 3 is a standalone product, independent of the earlier capstones: an AI Document Assistant — a "chat with your documents" tool where a user uploads PDFs, reports, or notes and gets grounded, cited answers instead of a generic chatbot guessing. It is deliberately the most widely-recommended, most teachable production RAG project in the industry — simple enough to explain to someone else end to end, real enough to actually demonstrate every stage of the AI & Machine Learning module, from embeddings through evaluation.
Topic Map
Everything in this module, at a glance
In Depth
Every topic, explained — with real-world industry context
Capstone Project 3 — Document Ingestion & Knowledge Base
Before any question can be answered, an uploaded document has to be turned into something an AI can actually search — this phase is that entire pipeline, end to end.
Core Build
- File upload — accepting PDF, DOCX & plain text
- Extracting raw text from each file format
- Chunking strategy — splitting long documents into retrieval-friendly pieces
- Generating embeddings for each chunk & storing them in a vector database
Product Features
- A document library — listing every file a user has uploaded
- Per-document metadata — title, page count, upload date
- Handling multi-page & multi-document uploads in one session
- Basic progress feedback while a large file is being processed
Production Polish
- Deduplicating identical or near-identical uploaded documents
- Re-processing a document cleanly if it's replaced/updated
- Handling malformed or unparseable files gracefully
- Per-user data isolation — one user never retrieves another user's documents
Real-World Industry Use Case
This is the exact ingestion pipeline behind every real "chat with your PDF" tool on the market — and per-user data isolation specifically is the same access-control principle named as a core production RAG concern, just applied to personal documents instead of enterprise ones.
Capstone Project 3 — AI Q&A Engine (RAG)
This is the reasoning core of the product — retrieving the right chunks of a user's own documents and turning them into a grounded, trustworthy answer.
Core Build
- A core RAG pipeline — retrieve relevant chunks, then generate an answer from them
- Prompt engineering — instructing the model to answer only from retrieved content
- Calling an LLM API (OpenAI/Anthropic) from the application layer
- Returning citations — which document & page an answer came from
Product Features
- Document summarization — a one-click summary of an uploaded file
- Multi-document questions — answering from more than one file at once
- Conversation memory — follow-up questions within the same session
- Clearly saying "I don't know" when the answer isn't in the documents
Production Polish
- Guardrails — preventing the model from answering confidently with no real source
- Comparing two retrieval strategies (e.g. keyword vs. hybrid vs. pure vector search)
- Cost-aware model routing for simple vs. complex questions
- Handling documents in multiple languages, if time allows
Real-World Industry Use Case
A well-built RAG assistant like this is consistently named as more impressive to hiring managers than a flashier but poorly-executed "original" idea — execution quality (clean grounding, honest "I don't know" answers, real citations) is what separates a hireable AI engineer's project from a tutorial clone.
Capstone Project 3 — Frontend (Next.js)
A clean, standalone Next.js application — its own product, not a feature bolted onto anything else.
Core Build
- An upload screen & document library (Next.js App Router)
- A chat interface for asking questions about uploaded documents
- Rendering citations as clickable links back to the source document/page
- Basic authentication — each user sees only their own documents
Product Features
- Streaming the AI's response token-by-token for a responsive feel
- A summary view alongside the chat for quick document overviews
- Clear loading & error states while a document is processing
- Simple document management — renaming or deleting an upload
Production Polish
- Accessibility — keyboard navigation & screen-reader support for the chat UI
- Responsive design for uploading & chatting from mobile
- Usage feedback — showing a user how many questions/documents they've used
- Polished empty & onboarding states for a first-time user
Real-World Industry Use Case
This is a direct, standalone application of the Frontend Development module — its own Next.js project from scratch, proving you can build a complete product UI independently, not just extend an existing one.
Capstone Project 3 — Backend & Cloud Deployment
A self-contained backend and infrastructure for this product alone — file handling, the AI pipeline, and hosting, all its own.
Core Build
- API endpoints for upload, processing & chat
- Object storage for uploaded files (S3 or equivalent)
- Secure server-side handling of LLM API keys — never exposed to the frontend
- A managed database for user accounts & document metadata
Product Features
- Background processing for large document uploads
- Streaming responses efficiently from backend to frontend
- Caching repeated questions to reduce redundant LLM calls
- Environment-based configuration across dev/staging/production
Production Polish
- A CI/CD pipeline for this project specifically — its own tests, lint & deploy steps
- Cost monitoring for both storage and LLM API usage, with budget alerts
- Auto-scaling the processing/chat service under real load
- A fallback path if the LLM provider has an outage
Real-World Industry Use Case
This is the Backend Development, Cloud Computing, and Version Control/DevOps modules applied as one complete, independent deployment — the full pipeline from Capstone Project 1 and 2, this time built around file storage and LLM-driven workloads instead of a course platform.
Capstone Project 3 — MLOps, Evaluation & Testing
An AI feature is never "done" the way a normal feature is — this phase is where you prove the assistant's answers are actually good, and stay good.
Core Build
- Building a small evaluation set — real questions with known-good expected answers, from real test documents
- Measuring retrieval quality — are the right chunks being found?
- Measuring groundedness — is every answer actually supported by a cited source?
- Basic unit & integration tests for the ingestion & RAG pipeline
Product Features
- LLM observability — tracing prompts, retrievals & responses end to end
- Testing the guardrails explicitly — confirming ungrounded answers are actually blocked
- A/B testing two prompt versions or chunking strategies against the eval set
- A simple dashboard for cost, latency & usage
Production Polish
- Detecting drift — does answer quality change as the underlying model updates?
- A regression-testing habit — re-running the eval set before every meaningful change
- An incident-response plan for "the assistant gave a wrong or unsupported answer"
- Presenting real eval metrics & a live demo as your capstone showcase to the cohort
Real-World Industry Use Case
This is the direct, applied continuation of the MLOps, Deployment & Responsible AI section of the AI & Machine Learning module together with the Testing & Debugging module — proving you can responsibly evaluate and maintain an AI feature, not just build one, which is exactly what separates a hireable AI engineer from someone who has only followed a tutorial.
Ready to learn this — live?
See how this module fits into the complete curriculum.