Hands-On Tutorial for Defining and Scoping Recommendation Systems
summary
Defines a five‑pillar “modern AI spec” and shows a live rewrite of a broken smart‑reply PRD into a measurable plan: behavior contract with good/bad canned examples and a ban on confidential names, evaluation ladder starting with a 500‑example golden offline set and an online guardrail that kills the feature if block rate rises by 0.3%, a hypothesis of >=10% reply‑time reduction and >=25% adoption, a V0 of hard‑coded top‑10 replies, allowed +300 ms latency, and a 5‑second kill switch.
AI-written summary · about
chapters (11)
- 0:00Why the 10-Page PRD Is Dead (and Build-First Isn’t Enough)
- 1:50The Problem With Vague Specs: ‘Helpful Replies’ and Other Non-Decisions
- 3:01Introducing the Modern AI Spec: The 5 Pillars That Make AI Products Work
- 4:05Pillar #2 Deep Dive: Behavior Contracts That Define ‘Good’ With Examples
- 4:56Pillar #3 Deep Dive: The Evaluation Ladder (From Offline Sets to Kill Switches)
- 5:59Live Demo: Rewriting a Broken Smart Reply PRD Using the 5-Pillar Template
- 6:34Opportunity Framing: Baselines, Hypotheses, and Targets That Actually Measure Impact
- 7:25Behavior Contract in Practice: Proxy Metrics, Guardrails, and Good vs Bad Replies
- 8:51V-Plan + Boundaries + Risks: V0 MVP, Latency Tradeoffs, and the Kill Switch
- 11:01Wrap-Up: The Spec as Leadership Tool (Alignment, Politics, and Skill-Building)
- 11:48Closing Thanks + Connect on LinkedIn
related talks
48mStop Fighting Your AI Engineering Prompts That Actually Work
19mOn Engineering AI Systems that Endure The Bitter Lesson
01h 01mIntroduction to Practical AI
41mRecommender Systems: Crafting Effective Personalised Experiences
40mSocially responsible AI: Practical tools for robust design
19mDon't Build Slop (4 Levels of AI Agent Maturity)
42mReasonable AI
43mAI Engineer Ethics: Shaping Responsible Generative AI