The AI Internship

July 3, 2026 · 8:00 AM PDT

Foundations

Agentic AI Engineering Bootcamp · Reliable reasoning

Dr Aki Wijesundara

Week 1

Dr Aki Wijesundara · The AI Internship01

The AI Internship

Meet your instructor

Dr Aki Wijesundara

Instructor · Agentic AI Engineering Bootcamp

PhD and Postdoc in Machine Learning, King's College London
Part of Google's AI Accelerator
AI Founder, SnapDrum
Senior AI Advisor to the United Nations
Previously: MIT, Oxford, Google
Co-founder and Lead Instructor, The AI Internship
(10,000+ students worldwide)

Dr Aki Wijesundara · The AI Internship02

The AI Internship

Logistics · plan, tools, schedule

Five weeks, one system

01

Reasoning

02

RAG

03

Agent

04

TRACE

05

Memory

Then a two-week build sprint + Demo Day

Set up before Week 1

Cursor or Claude Code · your coding agent
OpenAI + Anthropic API keys · with credit
Python 3.11+ · hello-world FastAPI runs locally

Live sessions · Fridays · 8:00 AM PDT

Jul 3Wk 1 Foundations

Jul 10Wk 2 RAG

Jul 17Wk 3 Agents

Jul 24Wk 4 TRACE

Jul 31Wk 5 Memory

Dr Aki Wijesundara · The AI Internship03

The AI Internship

Logistics · outcomes & community

What you leave with

A grounded, agentic Research Assistant at a live URL
Schema-validated /ask → RAG → agent → evals → memory
A TRACE eval suite tied to real failure modes
A Demo Day walkthrough of the full stack

A production system, not a tutorial repo.

Join the WhatsApp group

QR code to join Engineering cohort WhatsApp group

Scan to join the July Engineering cohort Home base for the bootcamp. Drop blockers, URLs, and wins here. Open WhatsApp invite →

Dr Aki Wijesundara · The AI Internship04

The AI Internship

Today's session

Week 1 agenda

~2 hours · 5 live demos · you ship /ask tonight

0

FrameReliable components vs one-off model calls

1

PlaygroundExperiment with prompts before writing code

Demo 1

2

FastAPIScaffold POST /ask with typed models

Demo 2

3

PromptingStructured outputs, few-shot, decomposition

Demo 3

4

GuardrailsValidate, retry once, reject cleanly

Demo 4

5

Token economicsCost, latency, and model selection

Demo 5

6

Ship itAssignment - a working /ask endpoint for your capstone

Dr Aki Wijesundara · The AI Internship05

The AI Internship

Segment 0 · Frame

A model call is easy.
A reliable component is engineering.

Anyone can get a plausible answer from a prompt once.
Production means: right shape, known cost, a plan for failure - every time.
This course is the reliability layer around a model you did not train.

A model call

prompt → answer

A reliable component

right shape

known cost

fails safely

Dr Aki Wijesundara · The AI Internship05

The AI Internship

The capstone spine

One system, built layer by layer

Today you build the reasoning core. Everything after adds to it.

Wk 1

Reasoning core

Wk 2

RAG

Wk 3

Agent

Wk 4

TRACE evals

Wk 5

Memory

→ Demo Day

Dr Aki Wijesundara · The AI Internship06

The AI Internship

What you ship today

`POST /ask` - question in, validated answer out

Schema-validated answer + tokens_used on every response.
Guardrail that rejects malformed output. A model choice you can defend.

Request body

"question": "…"

POST/ask

FastAPI endpoint

POST /ask

1Validate request schema
2Call chosen model
3Guardrail → validate response

200 OK

Response body

answer

tokens_used

Dr Aki Wijesundara · The AI Internship07

The AI Internship

Segment 1 · Playground

Start in the Playground, not in code

Experiment before you write software. Cheaper to learn a prompt here than in a redeploy loop.
System message sets role and rules. User message carries the task.

OpenAI Playground

System

You are a research assistant. Answer in 3 bullets max.

User

Summarise the key findings of this paper…

Run ▶

Open now platform.openai.com/playground →

Dr Aki Wijesundara · The AI Internship09

The AI Internship

The three dials that matter

Temperature · Max tokens · Token counter

Temperature

Low = consistent
High = varied

Max tokens

Ceiling on output - and on cost

Token counter

1,247

Cost as a number you watch move

Dr Aki Wijesundara · The AI Internship07

The AI Internship

Rule

Land the prompt here first.
Then take it to code.

Never debug prompt and plumbing at the same time.
Prompt behaviour becomes concrete against the real API, not theory.

Live demo 1 · ~5 min

Playground: same question at temp 0 vs 1. Lengthen the system prompt - watch the token counter. Land a working prompt to carry into code.

Dr Aki Wijesundara · The AI Internship08

The AI Internship

A prompt is not a system.
A system has an interface,
validation, and a failure plan.

You direct the agent. You read and judge every line it writes.

Dr Aki Wijesundara · The AI Internship09

The AI Internship

Segment 2 · FastAPI

Typed models = the contract

Shape in and out is defined and validated - not hoped for. Bad input rejected before the model call.

IN

Typed request

Pydantic validates first

→

GO

Handler

Model call lives here

→

OUT

Typed response

Shape defined, not hoped for

FastAPI first steps →

Dr Aki Wijesundara · The AI Internship10

The AI Internship

The /ask contract

Cost is a first-class field

Request

{ question: str }

Response

{ answer: str, tokens_used: int }

Live demo 2 · ~8 min

Scaffold FastAPI + POST /ask with Cursor / Claude Code. Read request model, model call, response model aloud. Hit it once - this is the spine everything attaches to.

Dr Aki Wijesundara · The AI Internship11

The AI Internship

Segment 3 · Prompting with rigor

Not clever wording - deterministic enough to build on

Structured outputs

JSON you can trust

Few-shot

Show the behaviour

Decomposition

Split the hard ask

Dr Aki Wijesundara · The AI Internship12

The AI Internship

Structured outputs

Chatbot → component

Chatbot

"Here's what I think… maybe…"

schema →

Component

{ answer, confidence, sources_needed }

OpenAI structured outputs →

Dr Aki Wijesundara · The AI Internship13

The AI Internship

Few-shot & decomposition

Few-shot

Show examples

Exact input-output behaviour you want. Often cheaper reliability than a longer instruction.

Decomposition

Break the hard ask

Steps the model handles reliably - not one prompt it fumbles.

Dr Aki Wijesundara · The AI Internship14

The AI Internship

Every technique has a price

Flag cost now.
Settle the bill in Segment 5.

Live demo 3 · ~6 min

Upgrade /ask to structured schema (answer, confidence, sources_needed). Show JSON in-shape. The endpoint becomes a component.

Dr Aki Wijesundara · The AI Internship15

The AI Internship

You taught the model your schema.
Now assume it sometimes won't.

Guardrails sit between model output and anything that acts on it.

Dr Aki Wijesundara · The AI Internship16

The AI Internship

Segment 4 · Guardrails

Validate · Retry · Never pass unchecked

model output

→

validate schema

→

downstream ✓

On fail: retry once → then reject cleanly

Pydantic validation →

Dr Aki Wijesundara · The AI Internship17

The AI Internship

One discipline, two directions

Structured output shapes the request. Guardrails defend the response.
Prompt injection through tool use returns in Week 3. Today: day-one fundamental.

Live demo 4 · ~5 min

Add validation + retry on malformed output. Force a failure - show the guardrail firing, not just sitting there.

Dr Aki Wijesundara · The AI Internship18

The AI Internship

Segment 5 · Token economics

Tokens = cost + latency

Treat the context window as a budget you spend on every call.

Context budget · per call

system

few-shot

question

output

× every call

Dr Aki Wijesundara · The AI Internship19

The AI Internship

Make the bill legible

Read `tokens_used` from your own endpoint

Connect it to the prompt choices you made in Segment 3.
Design with the budget in mind - don't discover the bill later.

Production best practices →

Dr Aki Wijesundara · The AI Internship20

The AI Internship

Model selection

Capability · Cost · Latency

Reasoning model

Capability

Cost

Latency

Fast model

Capability

Cost

Latency

Models overview →

Dr Aki Wijesundara · The AI Internship21

The AI Internship

Most systems use a mix

Strong where it matters.
Cheap everywhere else.

incoming task

→

router

hard → strong model

routine → cheap model

response

Live demo 5 · ~6 min

Same prompt, swap model. Watch tokens_used, latency, quality. End with one sentence: why this model for this endpoint.

Dr Aki Wijesundara · The AI Internship22

The AI Internship

Segment 6 · Ship it

Your assignment: ship `/ask` for your capstone

Not a toy - the reasoning core you ground in data next week.

Done when

Runs locally

Returns validated structured output

You can state per-call cost

Justified model choice

Dr Aki Wijesundara · The AI Internship23

The AI Internship

A model call is easy.
A reliable component is engineering.

Ship your /ask. Week 2 grounds it in your data.

Dr Aki Wijesundara · The AI Internship24

Foundations

Dr Aki Wijesundara

Week 1 agenda

A model call is easy.A reliable component is engineering.

One system, built layer by layer

POST /ask - question in, validated answer out

Start in the Playground, not in code

Temperature · Max tokens · Token counter

Land the prompt here first.Then take it to code.

Typed models = the contract

Cost is a first-class field

Not clever wording - deterministic enough to build on

Chatbot → component

Flag cost now.Settle the bill in Segment 5.

Validate · Retry · Never pass unchecked

Tokens = cost + latency

Read tokens_used from your own endpoint

Capability · Cost · Latency

Strong where it matters.Cheap everywhere else.

Your assignment: ship /ask for your capstone

A model call is easy.
A reliable component is engineering.

`POST /ask` - question in, validated answer out

Land the prompt here first.
Then take it to code.

Flag cost now.
Settle the bill in Segment 5.

Read `tokens_used` from your own endpoint

Strong where it matters.
Cheap everywhere else.

Your assignment: ship `/ask` for your capstone