FOR FRESH GRADS & ENGINEERS UNDER 2 YEARS

The First
AI Project
Kit
Pick one project. Ship it this weekend. Walk into your next
interview with something to talk about that isn't a to-do
app.

— 5 project blueprints, each with a built-in senior signal

— The exact Claude Code prompts for building and debugging

— A README structure written for a 30-second reviewer

— The ship checklist to run before you post anything

[Link]
LARA WEHBE @larawehbee_
READ THIS FIRST — 2 MINUTES

Why most junior portfolios fail

They prove you can follow a tutorial. They don't prove you can make decisions.

Hiring managers are not scanning your repo for "used the OpenAI API." They're scanning for evidence
that you thought about failure cases, cost, latency, and whether the thing actually works when the
input is ugly. That gap — between following steps and making calls — is the entire distance between
a junior portfolio and a hireable one.

This kit gives you five blueprints. Each has a senior signal: the one feature that makes a reviewer
stop scrolling. Build the boring version first, then add the signal. That order matters, and almost
everyone gets it backwards.

How to use this kit

1. Read the five blueprints. Pick the one closest to the job you actually want.
2. Fill in the Build Board so your scope is written down before you open the editor.
3. Work through the Prompt Pack in order.
4. Run the Ship Checklist before you post about it anywhere.

THREE RULES THAT MATTER MORE THAN YOUR CODE

• One project, finished. Three abandoned repos are worth less than one that works.

• Scope down twice. Whatever you think v1 is, cut it in half. Then cut it again.

• Measure something. One honest number in your README beats three paragraphs of
description.

About the fear: you've read that AI is closing the door on junior roles. What's actually happening is
that the bar moved from "can you write the code" to "can you judge whether the output is right."
Everything in this kit is designed to build that second skill — because that's the one that's still hiring.

The First AI Project Kit 2

STEP ONE

Pick your project

01 Document Q&A That Cites Its One weekend

Sources
PROVES You can build retrieval systems people can trust — the most requested
skill in enterprise AI right now.

STACK Python, FastAPI, a local vector store (Chroma or Qdrant), any

embedding model, any LLM.

V1 SCOPE Upload 10–20 PDFs. Ask a question. Get an answer with the exact source
chunks shown underneath it.

CUT FROM V1 User accounts, chat history, a pretty frontend, multi-format support.

THE SENIOR SIGNAL

Write 20 test questions with known correct answers. Measure how often the right chunk
actually gets retrieved. Put that number in your README. Almost no junior does this, and it's
the first thing an experienced reviewer looks for.

02 From Messy Documents to Validated Two evenings

JSON
PROVES You can do the unglamorous work companies actually pay for.

STACK Python, Pydantic, any LLM with structured output, a folder of real PDFs.

V1 SCOPE Feed in 30 invoices or CVs. Get back validated JSON on a fixed schema.
Anything that fails validation lands in a rejects folder with a reason
attached.

CUT FROM V1 A UI, a database, handling every document type on earth.

THE SENIOR SIGNAL

The retry loop. When the model returns something that fails schema validation, feed the
validation error back and retry once. Log your success rate before and after. That's a
production engineering mindset in twenty lines of code.

The First AI Project Kit 3

 03 Bilingual Meeting Notes One weekend

Agent
PROVES You can handle real-world messy input — specifically the code-
switching that breaks most off-the-shelf tools in this region.

STACK Whisper or Azure Speech, any LLM, Python.

V1 SCOPE Drop in audio from a meeting where people mix Arabic and English. Get
back a summary plus action items with owners.

CUT FROM V1 Live transcription, speaker diarization, calendar integration.

THE SENIOR SIGNAL

Test it on genuinely code-switched audio and document what breaks. Show a before/after
of your prompt or preprocessing fix. Every company in the Gulf has this problem and almost
no portfolio addresses it.

04 An Eval Harness for LLM Two evenings

Apps
PROVES You think about quality, not just output. This is the rarest thing in a junior
portfolio.

STACK Python, pytest, a CSV of test cases.

V1 SCOPE A small tool that runs prompts against a model, scores outputs against
expected results, prints a pass rate. Run it against two models or two
prompt versions and compare.

CUT FROM V1 A dashboard, a database, sophisticated scoring methods.

THE SENIOR SIGNAL

Wire it into GitHub Actions so it runs on every push. A passing CI badge on an LLM project
says more about your seniority than any bullet point on your CV.

The First AI Project Kit 4

 05 A Tool-Using Agent With One weekend

Guardrails
PROVES You can build agents that don't spiral — the difference between a demo
and a system.

STACK Python, any LLM with tool calling, the GitHub API.

V1 SCOPE An agent that reads open issues in a public repo, classifies them, and
proposes labels. It suggests. It does not write.

CUT FROM V1 Autonomous actions, memory, multi-agent anything.

THE SENIOR SIGNAL

A hard cap on tool-call loops plus a full log of every decision the agent made. Then write
one paragraph on what happened when you removed the cap. You'll have a real story for
interviews.

The First AI Project Kit 5

STEP TWO

The Build Board

Fill this in before you write a line of code. It's also the thing you'll reference every time
someone asks what you're working on.

FIELD YOUR ANSWER

Project chosen

The one-sentence pitch

Who this is for

"Done" means

Explicitly NOT in v1

My senior signal

The number I'll measure

Deadline

Phases

Finish each one before starting the next. The order is deliberate — polish before it works end to end is
the most common way weekends disappear.

PHASE DONE DATE

Write the spec — before any code

Scaffold one end-to-end path that runs

Make it survive bad input

Add the senior signal

The First AI Project Kit 6

 Measure and record the number

Write the README

Record a 60-second demo

Ship it publicly

The First AI Project Kit 7

STEP THREE

The Claude Code prompt pack

Use these in order. Most people skip straight to "build me X" and then spend the weekend
untangling code they don't understand.

01 Spec before code

I'm building [project] as a portfolio piece. Before writing any code, write a short spec:
the core user flow, the file structure, the data that moves between components, and the
three most likely failure points. Ask me anything you need to know first. Do not write
code yet.

02 Scaffold one working path

Build the smallest version that runs end to end using the spec above. One input, one
output, no error handling yet. I want to see it work before we make it good.

03 Understand what you shipped

Walk me through what you just built, file by file. For each design decision, explain what
the alternative was and why you didn't pick it.

This is the difference between having a project and being able to defend it in an interview.

04 Break it on purpose

Here are five ways a user could break this: [list them]. Show me what happens for each.
Then fix only the ones that would embarrass me in a live demo.

05 Debug properly

Here's the error and the full traceback: [paste]. Before fixing anything, tell me your top
two hypotheses for the cause and how you'd tell them apart. Then test the more likely one.

The First AI Project Kit 8

06 Write the README

Write a README for this project aimed at a hiring manager who has 30 seconds. Structure:
what it does, why it's non-trivial, the measured result, how to run it locally, what I'd
do with another week. No marketing language. Be specific about technical decisions.

07 Interview prep on your own repo

You're a senior engineer interviewing me about this codebase. Ask me eight questions,
starting easy and getting harder. Wait for my answer before asking the next one. Tell me
when an answer is weak.

Run this the night before any interview. It's the highest-leverage hour you'll spend.

The First AI Project Kit 9

STEP FOUR

The README that gets interviews

# Project Name

One sentence: what it does and who for.

## The problem
Two or three sentences. What breaks without this.

## What makes it non-trivial

The specific hard part. <-- this is the section reviewers read

## Results
The number you measured. Retrieval accuracy, success rate,
latency, cost per request — one honest number beats none.

## Stack
Short list. No badges.

## Run it locally
Three commands maximum. Test them on a clean clone.

## What I'd do next

Two or three items. <-- quietly the most important section

That last section tells a reviewer you understand your own project's limitations — which is the exact
quality the interview is trying to test for. Leaving it out reads as not knowing where the edges are.

The First AI Project Kit 10

STEP FIVE

Before you hit publish

Clone the repo fresh and follow your own setup instructions. Fix whatever broke.

No API keys in the repo — check the commit history, not just the current files.

The README contains a real measured number.

A 60-second screen recording exists. No narration needed.

Commit history shows real progress, not one commit named "final".

You can explain every file to a stranger.

Then write the CV bullet

Built [what] using [stack], achieving [number] on [how you measured it].

Developed an AI-powered application leveraging cutting-edge LLM technology.

The First AI Project Kit 11

The gap isn't
talent. It's a
finished project.
The person who got the job usually wasn't smarter than you and
often didn't have a better degree. They finished one thing and could
talk about it clearly.

You now have the blueprint, the prompts, and the checklist. The only
variable left is whether you open the editor tonight.

‫خبرني اكتر على شو عم تشتغل اليوم؟‬

Message me on Instagram and tell me which project you picked — or where you're
stuck. I read every message.

[Link]
LARA WEHBE · @LARAWEHBEE_