Session 8, second half · Week 4 · Thu 29 Oct 2026

Present it. Defend it. Ship it.

The last session. You present the system you designed, defend it under real questions, and walk out with something you couldn't build six weeks ago: a production LLM system that holds.

Ehsan Gazar
Production-Ready Systems with LLMs and Agents · session 8 of 8
The final 90 minutes

The run sheet.

0–8How the review runspresent + defend
8–65Design reviewseveryone presents + Q&A
65–78Full circleback to your Week-1 goals
78–88What to practise nextthe field keeps moving
88–90Send-offgo build
By the end of tonight

You'll leave able to…

1Walk a room through a production LLM design in five minutes.
2Defend each decision with a trade-off and a trigger, without flinching.
3Give specific, useful design feedback to a peer.
4Name what you'll practise next to stay sharp as the field moves.
The format

Five to present, five to defend.

This is a real design review, the kind that gets you promoted. You drive the room; the group is your panel. Warm, rigorous, and honest, we find the gaps because we want the design to survive.

flowchart LR
  P["Present
5 min walkthrough"] --> Q["Q&A
hard questions"]
  Q --> F["Feedback
specific, kind"]
  F --> N["Next presenter"]
  N --> P
  classDef ok fill:#D6F5E3,stroke:#1F2937,color:#0E1726;
  class P ok;
The five-minute walkthrough

Overview first, then the two decisions you're proudest of.

Don't narrate all seven sections. Show the one-page diagram, state the problem and the bar, then go deep on the two hardest calls you made. Reviewers reward depth on the risky decisions, not breadth.

1
The picture
system on a page + the bar it must hit
2
Two hard calls
the trade-off and the trigger for each
3
The risk you kept
what you're accepting, and why it's fine
What a strong opening sounds like

Sixty seconds, verbatim.

"This is a billing assistant for paid users. The bar: answer in under two seconds, and never state a policy that isn't in our docs, because getting that wrong is a legal problem, not a UX one. It's a workflow, not an agent, the steps are known, so I didn't buy autonomy I don't need. The two hardest calls were retrieval and refund safety. Let me take those…"

Problem, bar, shape, then straight into the two hard calls. No throat-clearing, no tour of the easy parts.

Under fire

Defend by naming the trade-off, not by flinching.

A hard question isn't an attack, it's the point. The staff answer isn't "you're right, I'll change it" or "no, I'm right." It's "I chose against that because X, and here's the trigger that would make me revisit." Same move as Session 1, now for a whole system.

Flinch
"oh, good point, I guess I'd change it" — you didn't decide, you guessed.
→ instead →
Defend
"I traded that for X on purpose; I'd revisit if Y" — you chose, and you can show it.
When it's your turn to critique

Good feedback is specific and load-bearing.

Useful

"Your retrieval has no permission filter, an injected doc could exfiltrate. Where does access get checked?"

✗ Useless

"Looks good." / "I would've done it differently." Nothing the presenter can act on.

Useful

Point at one decision, name the risk, ask the question that exposes it. One sharp note beats ten vague ones.

✗ Useless

Redesigning it for them. The goal is to strengthen their design, not substitute yours.

Back where we started

Remember the moment it broke?

In your pre-course reflection you named the moment an LLM feature met production and fell apart. Look at your design now. That failure has an answer in it, a boundary, a budget, a guardrail, an eval. That's the distance you covered.

Week 1 · you
"it worked in the demo, then everything broke at once"
→ six weeks →
Week 6 · you
"here's the system, and here's exactly how it holds"
Stay sharp · the field keeps moving

Four habits that outlast any model release.

1
Start simple, always
workflow before agent, smallest fuzzy job. The reflex that ages well.
2
Put a number on it
budgets and evals turn opinions into decisions. Do it every time.
3
Assume hostility + failure
untrusted input, flaky model. Design for both by default.
4
Feed the flywheel
every failure becomes a test. Compounding beats clever.
Project 5 · the capstone

System Design Document + Review, due today.

Your finished design document, plus the presentation you just gave (and the notes you took from the room). This is the artifact you take to work, the proof you can take an LLM feature from demo to production.

Include
the seven-section document, the one-page diagram, and your review takeaways.
Then
apply one decision from it to a real system at work this week.
Go build

Not another demo. A system that holds.

You came in able to wow a room. You leave able to survive real traffic, real cost, real failure, and a real reviewer. That's the whole difference, and it's yours now.

Thank you 🎉
Production-Ready Systems with LLMs and Agents · Ehsan Gazar