Madhu
Devi

AI integration in product design and architecture has become inevitable. It drives real gains in efficiency and productivity, but it also adds real complexity, and without careful governance, that complexity becomes risk no one sees coming until it's a problem: a customer complaint silently matched to the wrong internal requirement because two sentences happened to share a few generic words; an AI model quietly deprecated mid-build, still answering, just wrong, with no alarm going off. That's changing what's expected of product owners and managers: staying closer to the product, tracking deliverables at a far more granular level, and noticing the small things that used to be safe to skip. The same way a person double-checks their own work before sharing it, AI output needs that same habit built in as real governance, not blind trust in the first answer. This tour-booking operations prototype is an attempt to give that governance a concrete, working form: not a policy document, but a real product where every AI decision stays traceable, at zero extra manual cost to the team.

Built for a nonprofit that had run out of known funders and was stuck doing manual grant research by hand. A weekly AI search finds new opportunities, but the actual scoring is handled by deterministic logic, not the AI, so results stay explainable instead of feeling like a black box. Data ownership stays with the nonprofit itself, never a third party. Sign-in, match scoring, and a shared team tracker, all built end to end with Claude.ai and Claude Code as the build partners.

A wellness and recovery app for people who stand all day — restaurant workers, nurses, retail associates, and warehouse employees. A daily check-in captures shift data, a rule-based engine calculates a personalized recovery score, and an AI agent (LLM via Make.com) generates a warm shift story with emotional themes and one small recovery action. Supports English, Spanish, and Hindi. Built as a working demo with Lovable — full production version in progress with Next.js, Tailwind, and Make.com orchestration.

A peaceful daily space to feel, discover, and dream. Seven star angels, three curated cards — a daily gift, a place to dream about, and something to read or listen to. Built entirely with vanilla HTML, CSS, and JavaScript. No frameworks, no budget, no backend. Shipped with Claude as the sole build partner.

An interactive AI learning calendar — 120 concept cards plus 12 monthly milestones, arranged so the structure itself teaches. Four thematic parts rotate through the year: fundamentals, PM craft, ethics, and tools. Designed, curated, and shipped with Claude Project as the build partner.

Designed and shipped a personal website from zero to live using Claude AI as the sole development partner — no frontend skills, no budget, no templates. A case study in AI collaboration and conversation-driven building.

A practical knowledge base on managing annotation pipelines, RLHF workflows, and quality assurance for AI model development. Covers the full ADLC lifecycle, IRR arbitration, Red Teaming, and hands-on Labelbox practice.

An AI-powered speech evaluation companion for Toastmasters clubs. Built concept-to-deployment in a single sprint — real-time transcription, Pathways-aware evaluation criteria, and a privacy-first ephemeral architecture. Live experiment with a local club.