Level 4 Capstone: Ship Your AI Product

Build a Next.js 14 App Router customer support inbox app.
Stack: TypeScript, Tailwind, Supabase, Anthropic SDK.

Features:
1. Supabase table: tickets (id, subject, body, status, category, priority, created_at)
2. Supabase table: knowledge_base (id, content, metadata, embedding vector(1536))
3. POST /api/ingest — chunk and embed a text document into knowledge_base
4. POST /api/triage — given a ticket body, return {category, priority} using
   Claude claude-haiku-4-5 with structured JSON output
5. POST /api/draft-reply — RAG over knowledge_base, return a draft reply
   using claude-opus-4-5 with prompt caching on the system prompt
6. GET /api/tickets — list all tickets
7. UI: /inbox page with ticket list; click opens detail with:
   - Ticket body
   - AI triage badge (category + priority)
   - Suggested reply (editable textarea)
   - "Send" button (marks ticket resolved)

Add exponential backoff retry on all Claude calls.
Add per-request cost logging to console.

Build a Python multi-agent research assistant.

Architecture:
- orchestrator(question) -> list of 4 sub-questions
- researcher(sub_question) -> findings (bullet points)
  Use claude-haiku-4-5 for researchers (cheap + parallel)
- synthesiser(all_findings, original_question) -> 600-word report
  Use claude-opus-4-5 for synthesis
- critic(report) -> {"score": 1-10, "improvements": [...]}
  Loop max 3 times; stop if score >= 8 or "APPROVED" in output
- reviser(report, critique) -> improved report

Requirements:
- Use asyncio.gather() for parallel researcher calls
- Log token usage and cost per agent call to a CSV
- Add exponential backoff on all API calls
- CLI: python research.py "What are the tradeoffs of microservices?"
- Output: print final report + total cost summary

Eval: write 5 test questions with LLM-as-judge scoring
against ideal answers you write manually.

Build a Next.js 14 document intelligence app.
Stack: TypeScript, Tailwind, Supabase, Anthropic SDK.

Features:
1. /upload page: drag-and-drop zone accepting JPEG, PNG, PDF (max 10MB)
2. POST /api/extract:
   - Accept file as FormData
   - Convert to base64 (image) or base64 pdf document block
   - Send to claude-opus-4-5 with this system prompt:
     "Extract all data from this document. Return JSON:
      {type: 'receipt'|'invoice'|'form'|'contract'|'other',
       vendor: string|null, date: string|null,
       total: number|null, currency: string|null,
       line_items: [{description, quantity, unit_price, total}],
       notes: string}"
   - Validate JSON with try/catch; retry once on failure
3. Display result as editable table
4. "Export CSV" button for line_items
5. Supabase table: documents (id, filename, type, extracted_data, created_at)
6. /history page: list all processed documents

Add output safety check: if extracted total > 100000, flag for manual review.