Feasibility and Architecture

• The project, Shelf Scanner—an app that recommends books from shelf photos—began with a command-line MVP to test if cheap vision models could reliably read spines and make recommendations.1:24
• Latency, not cost, emerged as the primary bottleneck, with shelf-reading calls taking 10–20 seconds against a 15-second total experience goal.12:54
• Architecture decisions followed evidence: the speaker separated shelf-reading from recommendation, as different models—specifically Claude Sonnet and Gemini Flash for vision, and GPT 5.4 Mini for recommendations—excelled at distinct stages.10:53

Rigorous Workflow

• The speaker rejects "tech-first" development, instead documenting constraints and success metrics before writing code to ensure the AI serves the problem rather than vice-versa.1:56
• A spec-driven, human-in-the-loop workflow requires the coding assistant to draft a proposal for every feature; the human stops the process if the AI introduces an unmade decision.6:35
• Development includes versioning prompts, running deterministic software tests alongside AI evaluation sets, and stripping GPS metadata from user uploads to mitigate privacy risks.13:51

Career Framing

• The speaker argues that a portfolio project gains resume-worthy weight only if it evolves into a product with a real user base, transitioning it from a "learning artifact" to "professional experience."25:35
• The project avoids trendy techniques like RAG, agents, or fine-tuning, as the speaker asserts these add unnecessary complexity and cost without solving a verified problem.14:33