LLM project lifecycle
Scope → model choice → adapt (prompt/RAG/fine-tune) → evaluate → deploy → monitor — the GenAI-with-LLMs lifecycle for product teams.
The lifecycle (product view)
DeepLearning.AI’s Generative AI with LLMs organizes work as a lifecycle, not a single notebook. ShipAI uses the same spine:
flowchart LR
Scope[Scope & success metrics] --> Choose[Choose model / modality]
Choose --> Adapt[Adapt: prompt / RAG / FT]
Adapt --> Eval[Offline + online eval]
Eval --> Deploy[Deploy + gateway]
Deploy --> Monitor[Monitor / feedback]
Monitor --> Adapt
Skipping stages is how demos become outages: teams jump to a fancy agent before they can state a metric, or deploy without offline gates.
Scope before models
Write:
- User job-to-be-done
- Allowed failure modes
- Latency and cost budgets
- Data sensitivity class
- Who is accountable when the model is wrong
If you cannot state a metric (exact match, grounded citation rate, task success), you are not ready to pick a model.
| Scope artifact | Example |
|---|---|
| Success metric | ≥90% citation-valid on gold; p95 TTFT < 2s |
| Non-goals | No autonomous refunds |
| Data class | PII in tickets; redact before vendor |
| Fallback | Escalate to human queue |
Choose model / modality
Decisions at this stage:
- Closed API vs open-weight (Open-weight vs APIs)
- Text-only vs multimodal (Multimodal basics)
- Size / cost tier for interactive vs batch
- Tool-calling quality if you need agents
Ship rule: pick the weakest model that passes the eval bar — not the strongest model in the press release.
Adapt: three levers
| Lever | Use when | Cost to change |
|---|---|---|
| Prompt / tools | Behavior is linguistic; tools supply facts | Low |
| RAG | Answers need your corpus | Medium |
| Fine-tune | Style/format/domain language stable | Higher |
Order of operations for most products: prompt → RAG/tools → fine-tune. See examples on fine-tune vs prompt vs RAG and Advanced Fine-tuning LoRA/QLoRA.
flowchart TD
Metric[Metric failing] --> Q1{Missing private facts?}
Q1 -->|Yes| RAG[Improve retrieval / citations]
Q1 -->|No| Q2{Format / tool errors?}
Q2 -->|Yes| Struct[Schemas + prompts]
Q2 -->|No| Q3{Stable domain style?}
Q3 -->|Yes| FT[Consider LoRA]
Q3 -->|No| Data[Fix data / UX / scope]
Evaluate like a release gate
Offline golden sets → shadow traffic → canary. Never “demo vibe.”
| Stage | Purpose |
|---|---|
| Offline gold | Catch regressions in CI |
| Shadow | Compare new vs old on live traffic without user impact |
| Canary | Small % of users; watch metrics + cost |
| Rollback plan | One-click prompt/model revert |
Details: Evals fundamentals.
Deploy and monitor
- Gateway with auth, quotas, idempotency (Networking for AI apps)
- Streaming for interactive UX (Serving and streaming)
- Traces with prompt hash, model id, token counts
- Feedback loop: thumbs-down → gold candidates
- Privacy controls on logs (Privacy and data for AI)
Failure modes
| Failure | Stage skipped | Fix |
|---|---|---|
| Cool demo, no metric | Scope | Write exit criteria |
| Prompt spiral | Adapt | Time-box; change lever |
| Prod surprise | Eval | Offline gate + canary |
| Cost blow-up | Deploy/monitor | Budgets + routing |
| Silent quality drift | Monitor | Weekly gold refresh |
Tradeoffs
- Move fast on prompts — great early; hits a ceiling.
- Invest in RAG early — if facts live in docs.
- Fine-tune late — after prompts/RAG plateau and data is clean.
Glossary
| Term | Meaning |
|---|---|
| Adaptation | Changing behavior via prompt, data, or weights |
| Canary | Partial production rollout |
| Shadow | Score new system on live inputs without serving it |
| Exit criteria | Numeric bar to leave a lifecycle stage |
Micro-project
Write a one-pager lifecycle for your capstone idea with exit criteria per stage (scope → choose → adapt → eval → deploy → monitor).
Related guided path
Onboarding (rubric), Evals, guardrails, safety, Deploy, cost, latency, observability.