Skip to content

Projects

End-to-end projects connect theory to production. Each project is structured like a real engineering task.

Project Structure

Every project includes:

  1. Problem Statement — Business context and measurable success criteria
  2. Architecture — System design with Mermaid diagrams
  3. Implementation — Step-by-step build guide with full code
  4. Testing — Unit tests, integration tests, evaluation metrics
  5. Deployment — Docker, API, and cloud deployment instructions
  6. Optimisation — Profiling and performance improvements
  7. Interview Questions — Technical discussion questions about design decisions

Project Index

Project Volume Difficulty Estimated Time
P01: Spam Classifier Vol 3 Beginner 4 hours
P02: Image Classifier (ResNet fine-tune) Vol 4 Intermediate 8 hours
P03: Sentiment Analysis API Vol 5 Intermediate 8 hours
P04: Document Q&A with RAG Vol 7 Advanced 16 hours
P05: AI Research Assistant Agent Vol 8 Advanced 24 hours
P06: LLM Production Pipeline Vol 9 Expert 40 hours

P04: Document Q&A with RAG

Problem Statement

Build a production-ready Q&A system that answers questions over a corpus of PDF documents. The system must:

  • Achieve >80% answer relevancy on a held-out evaluation set
  • Return answers in under 2 seconds (p95)
  • Support up to 10,000 documents
  • Provide source citations with every answer

Architecture

graph LR
    User["User Query"] --> API["FastAPI"]
    API --> Embed["Embedding Model<br/>text-embedding-3-small"]
    Embed --> VDB["Vector Store<br/>(Qdrant)"]
    VDB --> Rerank["Re-ranker<br/>(cross-encoder)"]
    Rerank --> LLM["LLM<br/>(Claude claude-sonnet-5)"]
    LLM --> Response["Response + Citations"]

    PDFs["PDF Documents"] --> Chunker["Chunker"]
    Chunker --> Embed2["Embedding Model"]
    Embed2 --> VDB

Implementation

from pathlib import Path
import anthropic
from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct
from sentence_transformers import SentenceTransformer
import fitz  # PyMuPDF

EMBED_MODEL = "all-MiniLM-L6-v2"
COLLECTION = "documents"
CHUNK_SIZE = 512
CHUNK_OVERLAP = 64


def chunk_text(text: str, size: int = CHUNK_SIZE, overlap: int = CHUNK_OVERLAP) -> list[str]:
    words = text.split()
    chunks = []
    for i in range(0, len(words), size - overlap):
        chunks.append(" ".join(words[i : i + size]))
    return chunks


def ingest_pdf(path: Path, embedder: SentenceTransformer, client: QdrantClient) -> int:
    doc = fitz.open(path)
    text = " ".join(page.get_text() for page in doc)
    chunks = chunk_text(text)
    embeddings = embedder.encode(chunks, show_progress_bar=True)
    points = [
        PointStruct(id=i, vector=embeddings[i].tolist(), payload={"text": chunks[i], "source": str(path)})
        for i in range(len(chunks))
    ]
    client.upsert(collection_name=COLLECTION, points=points)
    return len(chunks)

def query(question: str, embedder: SentenceTransformer, client: QdrantClient) -> dict:
    q_vec = embedder.encode([question])[0].tolist()
    hits = client.search(collection_name=COLLECTION, query_vector=q_vec, limit=5)
    context = "\n\n".join(f"[{i+1}] {h.payload['text']}" for i, h in enumerate(hits))
    sources = [h.payload["source"] for h in hits]

    claude = anthropic.Anthropic()
    message = claude.messages.create(
        model="claude-sonnet-5",
        max_tokens=1024,
        messages=[{
            "role": "user",
            "content": f"Answer the question using only the context below.\n\nContext:\n{context}\n\nQuestion: {question}"
        }]
    )
    return {"answer": message.content[0].text, "sources": sources}

Evaluation

Measure retrieval quality separately from generation quality:

Metric Tool Target
Recall@5 Custom eval harness > 0.85
Answer Relevancy Ragas > 0.80
Faithfulness Ragas > 0.90
Latency p95 Prometheus < 2000ms