2026 · Cohort Level 1 · Chat-bot Edition Runs alongside Level 2 · Multimedia

From AI user, to AI builder — in four weeks.

A Caribbean AI + Python buildathon for high-schoolers across the ECCU. Students deconstruct, design, prototype, test, and pitch a human-led AI assistant of their own — grounded in a single rule for the whole programme.

Cohort
14 – 18 yrs
Duration
4 weeks · 20 days
Region
8 ECCU member countries · one currency union
The Route · 4 Weeks 20 days · 4 waypoints
W1 Deconstruct W2 Design W3 Build W4 Pitch N "you already know how to drive." STAGE MAP · CHAT-BOT EDITION
Route
Waypoint
ECCU territory
Not to scale
The One Rule · Autonomy ceiling · Safety anchor
The Bot is the GPS. / The Human is the Driver.
Repeated Day 1 · Day 7 · Day 14 · Day 20 · and every crash-test between them.
One currency union · Eight member countries 2025 camps ran in seven of eight · 2026 aims for all
§ 01 · The Journey

Four weeks. Four verbs. One prototype.

Each week is one movement in a compounding sequence. Miss a waypoint and the next one costs more; the map isn't shy about it.

Week 01
Deconstruct
Take one bot apart. Meet variables, strings, functions, and the A.R.T. classifier by Day 5.
Week 02
Design
Prompt-vs-Context split, few-shot, and persona anchoring — the real prompt engineering week.
Week 03
Build
Retrieval, chain-of-thought as visible planning, guardrails, and agent loops that actually close.
Week 04
Pitch
Testing + QA only. Confirmation register, cross-pod crash-test, 5-beat pitch, Ship-or-Wait.
§ 02 · Week Detail

Twenty days, each one carrying the last.

Every card names what's introduced, what it compounds on, and where the artefact lands in the camper's portfolio.

Week 01 · Days 1–501

From Curious User to AI Builder

Python basics land inside AI framing from the first hour. By Friday the pod has a working A.R.T. classifier and their first perceive/reason/act loop.
D1 · Deconstruct D2 · Design: Wardrobe D3 · Prototype D4 · Refine: Map + Rulebook D5 · Integrate + Agent Framing
Compounds on
Nothing — foundations. Introduces variables, dicts, functions, TONES register, and the six-component bot anatomy poster.
Week 02 · Days 6–1002

Prompt Engineering, Context & Safer Responses

Day 7 is the centrepiece: the Prompt-vs-Context split, few-shot template, and Persona Anchor — care-first, human-led, aligned to the one rule.
D6 · Register Recap D7 · Advanced Prompt Eng. D8 · Worked Banking Ex. D9 · PII Hygiene D10 · Camper Workbook
Compounds on
Week 1 functions and dict lookups; introduces the workbook: PROMPT · CONTEXT · three few-shot pairs · Persona Anchor.
Week 03 · Days 11–1503

Retrieval, Reasoning, Guardrails, Agent Loops

Day 4's rulebook was the toy. RAG is how it scales when the rulebook is fifty pages. Day 14 lands guardrails and structured JSON; Day 15 self-reflects on one showcase response.
D11 · Loop Anatomy D12 · Tool Use D13 · RAG + Selective D14 · Guardrails + JSON D15 · Self-Reflect + Recap
Compounds on
Days 1–10 combined. Chain-of-thought framed as visible planning notes — never expose private reasoning.
Week 04 · Days 16–2004

Testing, QA, Guardrails, and Pitch Readiness

Week 4 is testing and QA only. Advanced prompt-eng stays in Week 2, RAG in Week 3. Confirmation register, adversarial rotation, Ship-or-Wait, then the 5-beat pitch.
D16 · Mindset Re-anchor D17 · Confirm + Ship? D18 · Harden D19 · Mock Pitch + Debrief D20 · Premiere
Ship rule
Ship if every MUST-HAVE is signed, no Sev-1 open, all registers pass. Otherwise: Wait.
§ 03 · Day 01 · Vision-Cast

Before we touch the code.

The full Day 1 pulls the camper out of the "user" seat and into the "builder" one — mindset, two case studies, five-day map, and one word for the day.

You perceive how a bot works by taking one apart.

Day 1 is a mindset day, not a syntax day. Pods interrogate two camper-built bots — Miss Khalifa AI, the 2024 showcase winner, and Ancestra and The Guardian from 2025 — until they can explain what a bot actually does when you press enter, and where in that pipeline a human still has to sit.

"The Bot is the GPS. The Human is the Driver." Named on Day 1 · used through Day 20
01

AI Developer Mindset

Build · Test · Improve · Reflect. The four verbs that show up on every whiteboard, every day, for the next four weeks.

02

Miss Khalifa AI

Winner of the 2024 Chatbot Showcase — a camper-built sexual-health education chatbot on Python and LangChain. First specimen on the bench.

03

Ancestra and The Guardian

3rd runner-up, 2025 Chatbot Showcase — the second specimen. A placing bot and a winning bot fail in different places; pods find out where.

04

The 5-Day Map

Deconstruct → Design → Prototype & Test → Refine & Connect → Integrate & Pitch. Every day compounds on Day 1.

§ 04 · Two Levels · One Camp

Pick a track. New campers join Level 1; previous 2025 campers pick Level 2.

Level 1 is the four-week chat-bot buildathon for first-time campers. Level 2 is the multimedia edition for returning 2025 graduates — hosted on its own site.

Level 01 · This site

Chat-bot Edition

4 weeks · 20 days · Buildathon

The Level 1 pod deconstructs a bot on Day 1, designs prompts by Day 7, ships an agent loop by Day 14, and pitches on Day 20 with a claim ledger and a Ship-or-Wait call.

Duration
4 weeks · Jul 13 – Aug 9 2026
Age
14 – 18 years
Hours
Weekdays · 9:00 am – 4:00 pm
Cost
Free · ECCB & ICDF-supported
Level 02 · Multimedia edition

Coding, Music & AI Visual Design

17 days · From 09 July 2026 · For Level 1 graduates

Level 2 lives on its own master-hub, built by Enoete. This card is a doorway to it — we don't duplicate the content, we point at it and keep the reg link fresh.

Duration
17-day programme
Tracks
Code · 3D · Audio · AI Visual
Prereq
Level 1 graduate
Cost
Free · ECCB & ICDF-supported
§ 05 · The Client Board

Ten real enterprises. Ten real briefs. Your pod signs one.

These aren't invented scenarios. Ten Caribbean organisations filed chatbot discovery briefs for 2026 — banks, utilities, tax offices, social security, and the ports. Open a card to read the problem your pod could be pitching against by Day 20.

Sample Client · Fictional · Every studio trains here first
FP

FestPass ECCU Ltd.

Festival platform · All eight ECCU territories · "One pass. Every festival. Every island."

The flight simulator: every Studio learns to fly on FestPass before touching their real client's controls. Same skills, zero risk — and the client representatives are played by your facilitators.

  • Carnival season multiplies enquiries by twenty — and nine humans answer them all
  • "When does Carnival start?" has eight different true answers — the bot must ask WHERE first
  • The agentic red line: the bot can create a refund case, never approve one; open doors, never walk through them
  • The lost-child safeguard overrides everything — care first, humans alerted now, no promises
Sector

They want a bot that can read a water meter with its eyes.

Every school-age child in the federation with savings, shares — and a family full of questions.

The tax office answers the same twenty questions all day, every day.

A full-service consultancy in Roseau — 15+ years old, growing fast. Prospective clients don’t know which door to knock on.

Which fridge is cheaper to run? They want a bot that does the math.

Seventy years old, EC$2 billion strong — and the contact centre needs a brain.

Every channel, every hour — and in more languages than one.

Nineteen benefits. One bot to explain them all.

When the only tax officer is busy, who picks up?

Every plane, ship, and cruise passenger enters through them — and the phones never stop.

Plus five multimedia clients.

Five more enterprises briefed the Level 2 · Multimedia track — their board lives on the masterhub.

Level 2 masterhub ↗

Client Meeting Rooms — live.

Where the client sessions happen: the Day 3 interviews, the Day 8 mid-point check-in, the Day 17 confirmation working session. Find your studio’s room and join on time.

Enter the meeting rooms ↗
§ 06 · Caribbean case studies

Case studies — not demos.

GovChat Jamaica and SkyBot are discussion case studies: one public-sector, one private-sector example that campers analyse on Day 1, Day 7, and Day 17. The bots campers actually interrogate are a different thing — two showcase builds from past cohorts, on the bench below.

On the bench — the bots we interrogate

Day 1 Bot Probe · built by past cohorts
🏆 Winner · 2024 Chatbot Showcase

Miss Khalifa AI

The 2024 cohort's winning build — a sexual-health education chatbot on Python, LangChain, and React — comes back as this year's specimen. Day 1 pods probe it, map what works and what fails, and take its anatomy apart.

Chat live ↗ · GitHub ↗ · Drive ↗Camper-built · 2024
3rd runner-up · 2025 Chatbot Showcase

Ancestra and The Guardian

A 2025 showcase build on the bench beside the champion — proof that a placing bot and a winning bot fail in different places. Pods compare where each one holds and where each one breaks.

Chat live ↗ · GitHub ↗Camper-built · 2025
Public sector · Jamaica

GovChat Jamaica

A government assistant that answers "which office do I go to and what do I bring?" — quickly, in dialect, without pretending to be a case officer. We use it to teach scoping, hand-off, and the ceiling of a good agent.

Referenced · Day 1, Day 7Read the brief ↗
Private sector · Regional

Caribbean Airlines SkyBot

A customer-service bot that handles rebookings, baggage, and status. We use it to teach adversarial testing, care-first tone, and what should never be a bot's decision alone.

Referenced · Day 1, Day 17Read the brief ↗
§ 07 · Facilitator Room

Everything a lead needs, one door in.

The Facilitator Room holds every lesson plan, timetable, code artefact, image manifest, and safety brief in one place — kept in step with the curriculum so a facilitator never opens a stale slide by accident. Open access. No login.

Inside the room

Fifteen minutes to prep-ready.

  • 01 Day-by-day lesson plans5E structure · concepts · activities · code file · deliverable Open →
  • 02 Camper workbook & artefact templatesSix-Component Poster · Prompt/Context Split · Persona Anchor Card Open →
  • 03 Timetable & team formationECCU · SKN · per-island timing notes · pod roles Open →
  • 04 Image + slide manifestsWeeks 1–4 flag-report · what needs regenerating, what doesn't Open →
  • 05 Cross-week supplements & safety briefsClient-engagement plan · CHANGELOG · care-first appendix Open →
§ 08 · Repository

Every lesson, template, and file. Two shelves.

Campers get the decks, workbooks, templates, and code skeletons. Facilitators get everything — including the Kahoots, cheat-cards, and game kits that make the magic look effortless. No logins, no paywalls; files live on Google Drive and this page is the index.

Week
Type
0 resources on this shelf

Live from Google Drive: rows link straight into the shared AI & Python IDD, Curriculum and Lesson Plans folder. Week folders carry the complete materials; the Supplements folder holds the reference shelves.

§ 09 · The Toolbox

Free tools first. Paid only where it counts.

The 2026 stack prioritises free tools at every teaching moment — pods deploy without per-camper costs, and teens learn responsible provider selection from day one. Two providers anchor it: Google AI Studio + Gemini (free) and OpenAI GPT (the paid, industry-standard teaching example).

Free-first is a curriculum position, not a budget accident. Every skeleton runs a fallback chain — free provider first, paid alternative behind it — so campers learn that choosing a model is an engineering decision with a cost column.

Code & certify

Days 1–20
Trinket
Python in the browser — where Day 1's first lines happen. No install.
FREE
Deepnote
Cloud notebooks · the day-in, day-out coding home across all territories.
FREE
freeCodeCamp
The Python certification track — the daily cert block, Days 2–19.
FREE
GitHub
Pod repos and the road to a public URL.
FREE

The bot's brain

Day 5 onward
Google AI Studio
Prompt-engineering workbench — pods start here every day; the Day 7 & 17 room.
FREE
Google Gemini API
The programmatic brain inside every wk[N]_bot.py.
FREE TIER

Ship it

Days 6–20
Streamlit
The UI framework every bot wears from the Week 2 dashboard onward.
FREE
Streamlit Community Cloud
One-click public deploy on Day 18 — your bot gets a real URL.
FREE
Gradio
Alternative UI — the Week 4 stretch option.
FREE

Voice

Days 12–20
gTTS
Google text-to-speech — the zero-cost default voice. No robotic voices allowed.
FREE

See & protect

Day 11
Gemini Vision
Vision-language for reading documents — the Sense day.
FREE TIER
Pillow (PIL)
Python image handling and the redact_pii pipeline — privacy is a build step.
FREE

Remember (RAG)

Day 13+
sentence-transformers
Turns the rulebook into embeddings.
FREE
Chroma DB
The vector store behind the Day 13 RAG upgrade.
FREE

Design & pitch

Day 2 · Wk 4
Google Stitch
AI-native UI and wireframe drafting — Figma's free replacement.
FREE
Pomelli
Google's brand-content generator — Week 4 pitch collateral and Day 19 dress rehearsal.
FREE TIER

Data

Day 6 onward
pandas · numpy
Data frames and numerics behind the Week 2 dashboard KPIs.
FREE

Play & vote

Day 1 · Wk 4
Kahoot · Blooket · Gimkit
The Day 1 Bot Probe question kit — three game platforms, one showdown.
FREE TIER
Slido
Live Q&A and polling — from Day 1 straight through Week 4 pitch feedback.
FREE TIER

Home room

Days 1–20
Google Classroom
Assignments, gradebook, attendance — the camp's delivery platform.
FREE

Open & self-hosted — the sovereignty tier

Wk 4 stretch · advanced pods

Runners · the software

Ollama
One-command local LLMs — works with the internet off.
OPEN
LM Studio
The GUI runner — model marketplace, no command line needed.
OPEN
llama.cpp
Bare-metal inference — the Raspberry Pi showcase trick.
OPEN

Open-weight model families

Meta Llama 3.x / 4 Mistral · Mixtral Google Gemma 2 / 3 Qwen 2.5 / 3 Microsoft Phi 3 / 4 DeepSeek R1 / V3

Gemma runs free inside Google AI Studio — same family as Gemini, so Day 17 can switch between them live. Phi's 3.8B fits on a plain CPU: the model-constraints lesson in one download.

Hosted open inference

Groq
Sub-second Llama and Mixtral — generous free tier.
FREE TIER
Together AI
100+ open models behind one API — facilitator fallback.
PAY/USE

Day 8 · the concept — open weights exist  →  Day 17 · same prompt, Gemini ↔ Gemma, live  →  Day 18 stretch · Ollama offline: no key, no internet, still answering

"No one else owns your bot's brain when you self-host — and when hurricane season takes the internet, the bot still answers."

Security, guardrails & PII

Days 8–19 · the security lane

Protect the person

Presidio (Microsoft)
PII detection and anonymisation — the Day 11 redaction pipeline.
OPEN
Guardrails AI
Output validation and PII scrubbing — how the Day 14 Guardrails Wall gets implemented.
FREE TIER
Llama Guard / prompt-guard
Jailbreak and prompt-injection detection — an open-weights safety model.
OPEN
OpenAI Moderation API
Content-safety classifier for the Day 9 adversarial QA round.
FREE

Protect the keys

detect-secrets (Yelp)
Pre-commit hook that catches leaked API keys before they ship — every commit, Days 8+.
OPEN
GitGuardian
Continuous repo-level secret scanning across the pod repos.
FREE TIER

The key rule's enforcement arm: safe_call() keeps keys out of camper hands, and these two catch the accidents.

Protect the code

Bandit
Python security linter — Day 18 hardening sweep.
OPEN
Snyk
CVE scanning on requirements.txt — the dependencies get audited too.
FREE TIER
OWASP LLM Top 10
The reference framework — prompt injection, data poisoning, and friends. Cited in the Day 19 handover.
FREE

Day 8–9 · API security + adversarial QA  →  Day 11 · redaction pipeline  →  Day 14 · Guardrails Wall, six checks  →  Day 17 · Confirmation Register  →  Day 18 · harden + weakness log  →  Day 19 · handover security section

"All ten are free at camp scale. Security is the strongest free-first story in the whole stack."

Facilitator-managed · behind the curtain

The paid tier and the fallback chain.

  • OpenAI GPT + API — the paid, industry-standard teaching example; steps in when Gemini rate-limits hit. Day 8 onward.
  • Anthropic Claude API — reasoning-heavy tasks and Week 4 resilience testing.
  • HuggingFace Inference + Spaces — open-model fallback and alternate hosting.
  • OpenRouter — one key, many providers; the Week 4 resilience lesson.
  • OpenAI TTS · ElevenLabs · Play.ht — the neural voice upgrade path; premium Caribbean voice is a client-budget decision, not a default.
  • Full AI-Tools Inventory — the complete free-first ordering with day-by-day usage matrix lives in Playbooks & Supplements ↗.
"Campers never handle keys. Every skeleton routes through safe_call() with an env-var lookup — keys are camp-managed, spend-capped, and rotated when camp ends." The key rule · non-negotiable

Deprecated: pyttsx3 → gTTS chain  ·  Figma → Google Stitch

§ 10 · Safety & Judgement

The bot suggests. A human decides.

Care-first is a design principle, not a slogan. We name where it lives in the pipeline so pods know when to slow down.

What we mean when we say "care-first."

Any answer that touches money, health, family, immigration, or bereavement gets a slower path — a Persona Anchor pinned, PII redacted, a human name attached to the decision, and an escalation route that's actually used.

Autonomy ceiling

Bot suggests, human decides.

Named Day 1. Re-anchored Day 16. If the pod can't say who the human is, the artefact isn't ready.

PII hygiene

No real names in prompts.

Introduced Day 7 with the few-shot template. Redaction is a habit, not a rule they read once.

Guardrails wall

Six-check pre-response gate.

Built Day 14, used Day 18. The wall is a physical checklist on the studio wall, not a comment in code.

Ship-or-Wait

Signed register or wait.

Day 17 confirmation register. Client-facing claim ledger. If any Sev-1 is open, the pod waits — that's the whole rule.

§ 11 · Past Camps

Two cohorts in. The receipts are on YouTube.

The 2024 chatbot showcase, the 2025 regional showcase — nearly 400 campers across seven member countries — and the first episode of the 2026 season, already on air.

Short clips & recaps

For the bulletin · assemblies · parent evenings