For instructors

Master Timing Guide — all 6 sessions

Every deck has a live presenter timer built in — press S during the talk. It shows the per-slide budget, a cumulative target, a pace badge (on-time / behind / ahead), and speaker notes. This file is the printable version of the same data.

Delivery contract

Each session is a protected 120-minute instructional block. The lab is part of that block; breaks and lunch are outside it. The six blocks total 12 instructional hours. Day 1 runs S1 → S2 → S3; Day 2 runs S4 → S5 → S6. Confirm the room schedule separately so students receive recovery time between blocks.

The published slide budgets are targets, not permission to overrun. A lab-brief row includes the lab when stated. Slides marked ▸ are compressible; if a session is behind, cut those slides first. Do not consume the lab or capstone demo time.

Golden rule: protect the lab and the demo. If you’re behind at the lab-brief slide, compress the ▸ slides — never the hands-on work.

Session 1 — How Machines Learned to Talk

29 slides · core content ≈ 115.5 min; reserve ≈ 4.5 min for guided discussion, troubleshooting, and late pairs

#	Slide	Budget	Cumulative
1	How machines learned to talk	1 min	1:00
2	I sat in these seats — TCE CSE, class of 2009.	1 min	2:00	▸
3	12 hours. 6 things you’ll build.	1.5 min	3:30
4	Two kinds of AI: judges vs creators	3 min	6:30	▸
5	AI, ML, GenAI, LLM — who lives inside whom	2.5 min	9:00
6	ChatGPT is autocomplete at scale. That one idea goes surprisingly far.	1.5 min	10:30
7	Play the model: guess the next word	3.5 min	14:00
8	Those odds aren’t magic — you just count words.	2.5 min	16:30	▸
9	Watch the network think	3 min	19:30
10	Why the same question gives different answers	3 min	22:30
11	Wait — is it just searching a giant database?	2.5 min	25:00
12	What’s inside a model? Just knobs.	3 min	28:00
13	Training vs using: the cookbook rule	2.5 min	30:30
14	Models don’t read words. They read tokens.	2.5 min	33:00
15	You saw this machine on the pre-class page — here it is in one loop	3 min	36:00
16	If autocomplete can pass your exam, your exam was never testing understanding.	1.5 min	37:30	▸
17	The loop is old. The scale is new.	2 min	39:30	▸
18	A freshly trained model won’t answer you. It just keeps writing.	2.5 min	42:00
19	Finishing school, in three steps	2 min	44:00
20	Reasoning models: think longer, not just train bigger	2 min	46:00	▸
21	Why ChatGPT ≠ Gemini ≠ Claude	1.5 min	47:30	▸
22	Famous failures — you explain them	2.5 min	50:00
23	The context window: its entire working memory	2 min	52:00
24	Eight words you own now	2 min	54:00
25	What actually happens when you call an API	2.5 min	56:30
26	Your lab kit, and the rhythm	2 min	58:30
27	Lab 1: your first AI API call	50 min	108:30
28	Three things before the next session	5 min	113:30	▸
29	You’ve learned the trick. Next: you learn to drive it.	2 min	115:30	▸

Session 2 — Talking to AI, and Catching Its Lies

21 slides · core content ≈ 102.5 min; reserve ≈ 17.5 min for guided discussion, troubleshooting, and late pairs

#	Slide	Budget	Cumulative
1	Talking to AI, and catching its lies	1 min	1:00
2	Still true after the break?	2 min	3:00
3	Two artifacts in two hours	1.5 min	4:30
4	The prompt makeover: five upgrades	4 min	8:30
5	Anatomy of a prompt: six switches	2.5 min	11:00
6	Few-shot: show, don’t tell	2.5 min	13:30
7	Step-by-step beats straight-to-answer	2.5 min	16:00
8	Demand a format, or parse chaos forever	2 min	18:00	▸
9	The four classic prompt crimes	3 min	21:00
10	Your polished prompt still lies beautifully.	2 min	23:00
11	One of these is a confident lie	4.5 min	27:30
12	Hallucination is an expected failure mode. That does not make it acceptable.	1.5 min	29:00	▸
13	“How do you know it’s right?” You measure.	2 min	31:00
14	Watch an eval run	3.5 min	34:30
15	Same answer, three verdicts	3 min	37:30
16	Prompt A vs Prompt B: the arena	3.5 min	41:00
17	“It worked when I tried it” is not evidence.	1.5 min	42:30	▸
18	Eval-driven development: the loop	2.5 min	45:00
19	Six ideas you own now	2.5 min	47:30
20	Lab 2: the lie detector	52 min	99:30
21	You can now prove whether AI is right.	3 min	102:30

Session 3 — AI Beyond Text

17 slides · scheduled total ≈ 120 min

#	Slide	Budget	Cumulative
1	AI beyond text: eyes, ears and a paintbrush	1.5 min	1:30
2	Still true after the break?	2.5 min	4:00
3	Same loop. New kinds of tokens.	2 min	6:00
4	How a model reads a picture	5 min	11:00
5	It doesn’t just see. It reads.	3.5 min	14:30
6	Will it read it? Place your bets	5 min	19:30
7	Reading pictures is prediction. Generating them reverses noise.	2.5 min	22:00
8	Diffusion: a picture emerges from static	6 min	28:00
9	Image generation: magic with fine print	3.5 min	31:30	▸
10	Speech: useful, but not uniformly solved	4.5 min	36:00
11	Your mother’s voice is no longer proof of your mother.	1.5 min	37:30	▸
12	All of it is one API call	4 min	41:30
13	Where vision quietly fails	5 min	46:30
14	Vision ideas that belong to Madurai	3.5 min	50:00
15	Five ideas you own now	3 min	53:00
16	Lab 3: interrogate your photos	53 min	106:00
17	Day 1: you understand the machine. Day 2: you arm it.	14 min	120:00

Session 4 — Giving AI Your Own Knowledge

20 slides · core content ≈ 95.5 min; reserve ≈ 24.5 min for guided discussion, troubleshooting, and late pairs

#	Slide	Budget	Cumulative
1	Giving AI your knowledge	1 min	1:00
2	Did it survive the night?	2 min	3:00
3	Watch it not know — loudly	3 min	6:00
4	“Just paste all my notes!” — let’s price that	2.5 min	8:30
5	Don’t send everything. Send the right 3 paragraphs.	1.5 min	10:00
6	Keyword search misses meaning	2.5 min	12:30
7	Embeddings, now for whole paragraphs	2.5 min	15:00
8	The search playground	3.5 min	18:30
9	Chunking: how you cut the book	3 min	21:30
10	Vector databases, in plain words	2 min	23:30	▸
11	RAG, end to end	4 min	27:30
12	The grounded prompt template	2.5 min	30:00
13	Where RAG breaks in the wild	3 min	33:00
14	RAG vs paste-it-all vs fine-tuning	2 min	35:00	▸
15	You’ve been using RAG all along	1.5 min	36:30	▸
16	Your final-year project needs a user, a baseline, and a number.	1.5 min	38:00	▸
17	Lab architecture: two pipelines	2 min	40:00
18	Six ideas you own now	2.5 min	42:30
19	Lab 4: chat with YOUR notes	50 min	92:30
20	Your AI now knows what you know.	3 min	95:30

Session 5 — Making AI Do Things

23 slides · scheduled total ≈ 120 min

#	Slide	Budget	Cumulative
1	Making AI do things	1 min	1:00
2	Still true after lunch?	3 min	4:00
3	Your RAG app is a brain in a jar	2 min	6:00
4	The model never executes anything. It writes requests. Your code executes.	2 min	8:00
5	One tool call, step by step	5 min	13:00
6	Function declarations: the description IS the prompt	3 min	16:00
7	MCP: USB-C for tools	1.5 min	17:30	▸
8	Which tool should it call?	4 min	21:30
9	The agent loop: while it wants tools, feed it	3 min	24:30
10	Five ways tool use goes sideways	3.5 min	28:00
11	Most problems don’t need an agent. They need a boring workflow.	2.5 min	30:30
12	Why long agent chains collapse	5 min	35:30
13	Multi-agent: more agents, more compounding	2 min	37:30	▸
14	Most production “AI agents” are a while-loop in a trench coat.	1.5 min	39:00	▸
15	Workflow or agent? One question decides	2.5 min	41:30
16	The escalation ladder: cheapest fix first	3.5 min	45:00
17	Right approach? Defend your vote	4 min	49:00
18	API models vs models you own	2 min	51:00	▸
19	Ollama: a model in your pocket	4 min	55:00
20	Local vs API: an engineering trade, not a religion	1.5 min	56:30	▸
21	Six ideas you own now	3 min	59:30
22	Lab 5: give it hands	50 min	109:30
23	Your AI knows, sees, and now acts.	10.5 min	120:00

Session 6 — Breaking, Securing, Shipping

21 slides · scheduled total ≈ 120 min

#	Slide	Budget	Cumulative
1	Breaking it, securing it, shipping it	1 min	1:00
2	Everything, in 90 seconds	2 min	3:00
3	For five sessions you built. Now: think like an attacker.	1.5 min	4:30
4	Prompt injection: data becomes commands	3 min	7:30
5	Indirect injection: the document attacks your RAG	3 min	10:30
6	Jailbreaks & leaks: the two quieter doors	2.5 min	13:00
7	Defense in depth: no single fix, so layer them	3.5 min	16:30
8	Keep a human on anything that bites	2 min	18:30
9	There is no perfectly secure AI agent. Build one whose blast radius you can survive.	1.5 min	20:00	▸
10	It works in Colab. Now make the assumptions visible.	1.5 min	21:30
11	Cost: every token is a coin	3 min	24:30
12	Fast, resilient, and observable — speed, reliability, observability	2.5 min	27:00
13	UX patterns for honest AI products	1.5 min	28:30	▸
14	Responsible AI: the four questions auditors now ask	2 min	30:30	▸
15	Your users decide whose law you’re under	2.5 min	33:00	▸
16	The ship-it checklist	2.5 min	35:30
17	Capstone: the sprint & the demo	2 min	37:30
18	How to demo without dying	2 min	39:30
19	Six sessions, one throughline	76 min	115:30
20	Where to go from here	2.5 min	118:00	▸
21	You came as users. You leave as builders.	2 min	120:00