#The Panel Presentation

The round that decides field loops. You get a scenario in advance, present to a panel role-playing customer stakeholders, and they interrupt. Candidates who are otherwise strong fail here, and candidates who are technically lighter sometimes win on this alone.

Typical brief: "You are the architect meeting a customer for the first time. Here is their situation. Present your approach." Sometimes it is "teach us a technical topic as if to a customer." Both are graded on the same things.

#1. What they are actually scoring

The single most common failure is presenting a solution to a problem you never asked about. It is also the easiest to fix.

#2. The structure

10–14 slides. Never more. Assume you will lose a third of your time to interruption — that is the design of the exercise, not bad luck.

Optional appendix slides for anticipated deep-dives — have them, don't present them. Pulling up an appendix slide in response to a hard question is one of the strongest moves available to you.

#3. Slide 3 is the whole game

Most candidates put discovery in a bullet list and move on. Do this instead:

"Before I show you anything, there are five things that would change what I'd recommend, and I want to flag them so you can correct me as I go.

One — freshness. I've assumed hourly is enough for the reporting, and near-real-time only for the operational alerting. If the trading desk needs sub-minute, the ingestion design changes materially. Two — residency. I've assumed EU data can be processed in EU regions and aggregates can move. If raw data can't leave a country, we need a metastore per country and the reporting design changes. Three — who operates this in twelve months. I've assumed your platform team, six people. If it's two, I'd design for far less operational surface — more serverless, more managed. Four — the Teradata contract date. If it's 18 months, the sequencing works. If it's nine, we're having a different conversation. Five — is the AI mandate funded, or is it aspiration? That changes whether I build for it now or leave the door open.

Where am I wrong?"

That last question is the move. It invites the panel into the conversation, which is the entire job. It also gives you their priorities before you commit, so the rest of your presentation can adapt — and adapting visibly is exactly what scores on coachability.

#4. Delivery mechanics

Open standing up and with energy, even remotely. The first fifteen seconds set how hard the panel works to stay engaged.

Narrate diagrams in a fixed order. Governance → ingest → storage layers → consumption → operations. Never point at a box and say "and this is Kafka." Say why the box exists.

Handle interruptions with a three-beat move:

1. Stop mid-sentence. Immediately. Never finish over an interviewer.
2. Answer the question actually asked, at the altitude it was asked.
3. Return explicitly: "— so coming back to the ingestion layer…"

Read the room's altitude. If the "CFO" asks a question, answer in money and risk. If the "lead engineer" asks, answer in mechanism. Answering a CFO with shuffle partitions is a scored failure.

When they role-play disengagement — and some panels deliberately do, sitting flat and not reacting — do not let it drain you. Name it lightly and keep going:

"I'd normally draw this out with questions here — I'll keep moving and leave time at the end."

Time. Land your core message by 60% of the slot. If you're at slide 5 with ten minutes left, say so and skip: "I'm going to jump to the sequencing and the risks, because those matter more than the rest of the architecture detail — and I'll leave the detail in the appendix." Visible time management is itself a competency being scored.

#5. If the brief is "teach us a topic"

Pick something you can go genuinely deep on and explain simply. Good candidates: the Delta transaction log and why it gives you ACID on object storage · why data layout dominates query cost · governing GenAI at enterprise scale · the real economics of serverless.

Structure it as: why this matters to a business → the mechanism → a worked example → where it breaks → what you'd do about it. Finish with the trade-off, always.

Do not pick the topic you know best if you cannot make it matter to a customer. The round is about teaching, not about knowing.

#6. Preparing yours

Build it on Day 8 of the sprint. Then:

1. Deliver to a wall, timed. It will be too long. Cut 30% — cut content, not pace.
2. Deliver to a person who is instructed to interrupt three times, including once with something hostile and once with something off-topic.
3. Deliver to someone non-technical. If they can't say back what the customer gets, slide 4 is wrong.
4. Prepare the appendix for the five deep-dives you most expect.
5. Prepare for the technology to fail. Have a PDF. Be able to deliver it with no slides at all — and say so if it happens: "I'll talk you through it and share the deck after." Handling that gracefully is a free competency demonstration.

#7. The ten questions you will be asked

Rehearse a 60-second answer for each against your own deck:

1. "Why not just use Snowflake / what we already have?"
2. "What's this going to cost?"
3. "How long before we see anything?"
4. "We don't have the skills for this. Then what?"
5. "What's the biggest risk in what you've proposed?"
6. "What would you do differently with half the budget?"
7. "Our last platform programme failed. Why is this different?"
8. "Can you prove to our regulator that this is secure?"
9. "What have you left out?"
10. "If you were us, what would worry you about hiring you?"

Question 9 is a gift. Answer it honestly and you demonstrate judgement, self-awareness and scope discipline in one move: "Three things deliberately — real-time serving, because I don't yet know if you need it; the ML platform, because governed data has to come first; and the migration of the legacy reports, because I'd want to know how many are actually used before committing. Any of those could be wrong, and the discovery in phase one is designed to tell us."

#8. The close

Land one sentence they will repeat to each other after you leave. Not a summary — a thesis.

"If you take one thing from this: the constraint on your AI roadmap isn't models or compute, it's governed context — so the sequencing I've proposed spends the first ten weeks making your data governable, because everything after that goes three times faster."

Write yours. Say it last. Then stop talking.