TAI Labs
TAI Lightning Lesson · Free

Design AI
projects
recruiters
care about

Build for the deep dive. Not just the scan.
Session
Live · one sitting
You bring
One project idea worth the deep dive
Led by
Aki WijesundaraAki Wijesundara
Manu JayawardanaManu Jayawardana
TAI Labs
The problem

Your portfolio gets scanned. Then it gets interrogated.

Most candidates build for the scan. Clean landing page. Nice demo video. Then the technical deep dive hits and the wheels come off. You are asked why not X, three times, and none of your answers are ready.

The point of the next thirty: a framework called SCOPE, five README headings that pre-answer most of the interview, and a template you can apply to any project you have.
TAI Labs
By the end of this session

SCOPE, an eval set, and a README recruiters read all the way through.

The receipt: the SCOPE framework, a 30-case eval structure with baseline, and a five-heading README template you can apply to any project you already have this weekend.
01
Learn the SCOPE letters. Situation. Constraints. Options. Proof. Edges. Say them out loud.
02
Focus on C, O, P. Constraints, Options, Proof. That is where offers are lost.
03
Ship an eval set + baseline. The two artefacts a technical reviewer respects the most.
04
Rewrite your README. Five headings. Answers ready before the interviewer arrives.
TAI Labs
Hosts

Who we are.

Dr. Aki Wijesundara

PhD in Machine Learning. Google AI Accelerator alum. Co-founder, SnapDrum. AI Advisor to the UN.

Manu Jayawardana

Exited AI founder. CEO, TAI Labs. Ten thousand plus engineers and PMs trained across our programmes.

TAI Labs
The interview

What gets checked. And when.

Stage
Length
What they check
Portfolio scan
Brief
Can I tell what this does.
HM review
Short
Did you know the real constraints.
Technical deep dive
Long
Why not X. How do you know it works.
Behavioural
In passing
Can you name a limitation.
Candidates optimise for row one and get eliminated in row three. The technical deep dive is where the offer is decided. Design for that.
TAI Labs
The gap

Demo is not judgement.

Row 1 rewards

A clean demo.

Video that plays. Landing page that loads. Buttons that click. Recruiters can tell what it does. This is table stakes.

Row 3 rewards

Documented judgement.

Why this approach and not the two obvious alternatives. What breaks it. How you know it works. The trade you made and what you gave up. This is what wins offers.

They are not the same project. A row-one project can be built in a weekend and read in the elevator. A row-three project takes deliberate design decisions you can defend under pressure.
TAI Labs
The framework

SCOPE. Five letters. Five answers ready.

S
Situation. Who is the user, and what does the current process cost. One paragraph. Facts only.
C
Constraints. Which two of data, cost, latency, failure did you optimise. Name the trade explicitly.
O
Options. What you rejected, and why. Three alternatives, three reasons.
P
Proof. A baseline and an eval set. Numbers, not vibes.
E
Edges. What breaks it. Say it before the interviewer says it.
The rule of the session

If your README has these five headings, you have pre-answered most of the interview.

Every hard question the technical deep dive throws at you is one of the five letters. Situation. Constraints. Options. Proof. Edges. If your README already answered it, the interviewer moves on. If it did not, they push.

TAI Labs
Focus today

C, O, and P. Where offers are lost.

S

Situation

Easy. Who and what cost.

C

Constraints

Where offers are lost.

O

Options

Where offers are lost.

P

Proof

Where offers are lost.

E

Edges

Easy. What breaks it.

S and E are the easy two. Say the trade out loud so it reads as deliberate, not accidental. The three in the middle are where the offer decision happens.
TAI Labs
C · Constraints

Same domain. Two projects.

Weak

A support chatbot over our help docs.

No constraint named. No trade visible. Vague scope. A reviewer can ask ten questions and each one is a stumble.

Strong

A ticket triage agent that routes by category and escalates below a confidence threshold, measured against the current human baseline.

Constraint named (accuracy plus failure behaviour). Trade visible (escalation over guessing). Baseline stated (human). A reviewer knows what to ask.

Which one would you rather be asked about in the deep dive? Pick the one that gives you room to say something specific.
TAI Labs
C · Constraints

The four dials. Pick two.

01

Data quality

What the model can trust. Cleanliness, coverage, freshness.

Dial 1
02

Cost per request

What you can afford to run. Tokens, calls, infra.

Dial 2
03

Latency budget

How fast it must answer. p50 and p95 the user sees.

Dial 3
04

Failure behaviour

What happens when it is wrong. Escalate, refuse, guess.

Dial 4
You cannot optimise all four. Name the two you chose. Name what you gave up. That sentence is what a technical reviewer wants to hear on the deep dive.
TAI Labs
Halfway through

We stop. We hear the room. We keep going.

What we do
  • Post one project you have on your portfolio right now, in chat.
  • One line. Name the constraint you actually optimised.
  • Three or four get read aloud.
  • Then straight into Options and Proof.
What this is for

Most projects don't name a constraint.

Which is why the deep dive breaks them. If you can name your two dials without hesitating, you have already passed the hard part.

If you can't, the next two blocks are exactly what you need.

The claim under all of this: if you cannot name what breaks your project, it is not a project yet. It is a demo.
TAI Labs
C · the test

If you cannot name what breaks it, it is not a project yet.

This is the sentence that separates a demo from a project. Every project a reviewer respects has an answer to "what breaks it." If yours doesn't, the reviewer's next question is "so what have you actually shipped."

TAI Labs
O · Options

"Why not just use X?"

Asked three times. In three different ways. Write your answers down before the interview exists.
Q1

Why not fine-tune?

Cost, data-hunger, brittleness of the retrain loop. Say it before they say it.

Common ask
Q2

Why not a smaller model?

Where accuracy or reliability drops on your specific eval. With numbers.

Common ask
Q3

Why not rules?

Coverage of the long tail. Point at a case a rule cannot enumerate.

Common ask
The pattern: three prepared answers turn a jumbled section of the interview into three sentences that show you thought about it. That is the whole difference between a reviewer nodding and pushing.
TAI Labs
P · Proof

A thirty-case eval set. In one working stretch.

20

Normal cases

Real usage. Actual inputs the system will see. Not the ones that make the demo look good.

The floor
10

Expected fails

Cases you expect to break. Adversarial, ambiguous, or out-of-scope. Say why you expect them to break.

The edges
Score

Write the number down

Pass rate on normal. Pass rate on expected fails. If the second one is high, either your fails are not hard enough or your system is genuinely strong. A reviewer respects the framing either way.

The proof
An eval set beats a demo video. Reviewers can watch a demo. They cannot watch a demo prove itself against thirty different inputs the way an eval set can.
TAI Labs
P · Proof

No baseline. No result.

Means nothing

"92% accuracy"

Ninety-two percent of what. Compared to what. On which cases. A number alone tells a reviewer nothing about whether the system is worth shipping.

Means something

"92% against a human doing it manually."

Now there is a comparison. Now a reviewer knows what you traded for what. Now the number is a claim about the world, not a decoration.

Always ship the comparison. The baseline can be a human, a rule-based system, an off-the-shelf model, the previous version. Any comparison beats an isolated percentage.
TAI Labs
The artefact

The README template. Copy this today.

Five headings
## Situation ## Constraints ## Options considered ## Proof ## Edges
What goes under each
  • Situation. User + current cost, in one paragraph.
  • Constraints. Two dials you chose + what you gave up.
  • Options considered. Three alternatives + why not each.
  • Proof. Baseline + eval set + numbers.
  • Edges. Three things that break it + how you would know.
Print these headings. Paste them into every project README you have. A README a reviewer reads all the way through is worth more than a demo they watch once.
TAI Labs
What you do this week

Rewrite one README. Five headings. This week.

01
Pick a project with a real constraint. The one you would rather be interrogated about. If none of your projects have a constraint you can name, pick the closest and add one.
02
Document the option you rejected. Three alternatives, three reasons. Say them out loud until they sound natural. These are the three questions that show up in every deep dive.
03
Ship a baseline and thirty eval cases. Twenty normal, ten expected fails. Score them. Write the numbers down. One working stretch. That is all it takes.
04
Update the README with the five headings. Copy them from this deck. Fill each with one paragraph. A reviewer who reads all five is a reviewer who is now selling you internally.
The receipt: a GitHub link to the updated README. Send it. We open Week 1 of the Agentic AI Engineering Bootcamp by reviewing three of these together.
TAI Labs
The habit to take with you

Before you ship a project, write what breaks it.

If you cannot fill in the Edges section, the project is not ready to be interviewed about. The candidates who get offers at frontier labs are the ones who describe their own project's failure modes before the interviewer can. That posture is the whole difference.

Next

Agentic AI Engineering Bootcamp

Nine weeks. Ship production agents with evals, memory, and a portfolio that survives the deep dive. Certificate on completion. Cohorts start monthly.

Where this leads
Now

Questions in chat

Post the one project you are about to update. One line. Aki or Manu will read a few aloud and name the constraint.

Open floor
Later

Your README

Send us the GitHub link once the five headings are filled in. We reply with an audio review before the bootcamp starts.

The receipt
01 / 19