Build ONE of these well (better than three half-done)
Project A • Best fit
AI Ticket Classifier Pipeline
Webhook/API ingest → queue → LLM classify to JSON → store in SQLite/Postgres → simple report page. Golden-set eval script included.
APIs LLM DB Webhooks
Project B
Invoice / Document Extractor
Upload text/PDF text → structured JSON extraction with validation → CSV export. Shows prompt engineering + schema discipline.
Prompting JSON schema
Project C
n8n + Code Automation
n8n workflow (form → AI classify → Google Sheet/DB → Slack/email) + one step done in code. Shows low-code + code judgement.
n8n Integrations
Project D
RAG FAQ Assistant
Embed a small docs set → retrieve → answer with citations. Even 20 documents is enough to demo chunking & retrieval clearly.
RAG Embeddings
48-hour build plan (Project A)
| Block | Deliverable |
|---|---|
| Hours 0–4 | Repo, README skeleton, API endpoint + webhook receiver, .env.example |
| Hours 4–10 | LLM classify with JSON schema + validation + retry |
| Hours 10–16 | DB storage, dedupe by event id, queue/background processing |
| Hours 16–24 | Golden set (30 examples) + eval script printing accuracy |
| Day 2 | Report page, error handling, SOP doc, demo video/GIF, polish README |
README that gets you shortlisted
- What it does (2 lines) + architecture diagram or flow list
- How to run it in under 5 commands; .env.example included
- Design decisions & trade-offs (model choice, dedupe, retries)
- Eval results table — accuracy on your golden set
- What you'd improve with more time (shows senior thinking)
Hands-on labs (90 min each)
- Webhook lab: Receive, verify signature (HMAC), dedupe, return 200 fast, process async.
- Rate-limit lab: Call a public API in a loop; implement backoff + caching.
- Prompt lab: Same task, 3 prompt versions; score all on 20 examples.
- AI-coding lab: Add a feature with Claude Code/Codex; keep the diff + tests to show in the interview.
via API / webhook"] --> API["API endpoint
validate + dedupe"] API --> Q["Queue"] Q --> LLM["LLM classify
temperature 0, JSON schema"] LLM --> VAL{"Valid +
confident?"} VAL -->|Yes| DB[(SQLite / Postgres)] VAL -->|No| HQ["Human queue /
retry once"] HQ --> DB DB --> REP["Report page
counts by category"] LLM -.-> EV["Eval script
golden set accuracy"]
Test yourself
Module 5 Test — Projects & Labs (self-check)
Tap an answer — explanation appears instantly. Pass mark: 70%. Retake until you pass.
One strong portfolio project beats three half-done because…
Your README must contain…
Golden set is for…
0 / 3 correct