NHS Decision Lab

Inspriation I have spent more than four decades in sales, homebuilding, mortgage lending, and real estate education. During that time, I have observed that professional education often teaches information and then tests recall. But successful real estate agents need something harder to teach: judgment that benefits the client. Fifteen years ago, I began creating professional education courses for real estate agents focused on new-home construction and buyer representation in new-home transactions. A few years ago, I created the New Home Specialist curriculum to help real estate agents better represent buyers considering new construction. OpenAI Build Week gave me the opportunity to explore whether Codex and GPT-5.6 could transform that traditional curriculum into an interactive learning experience where agents practice responding to realistic client conversations and receive immediate, evidence-based coaching. I am not a software developer. My curiosity about learning how to use AI led me to enter Build Week as the product owner, subject-matter expert, and instructional designer behind an idea for improving professional education. Codex became the technical collaborator that helped turn my professional knowledge and approved course materials into a functioning prototype.

What it does NHS Decision Lab is an interactive learning simulation that allows real estate professionals to practice new-home buyer-representation decisions in a realistic client scenario. GPT-5.6 evaluates the learner’s response across defined professional-judgment criteria and provides personalized coaching, evidence from the response, a recommended improvement, and a stronger example to study.

How I built it I began with materials from a new New Home Specialist course I am creating, including the scenario, characters, instructional principles, and professional standards. I worked with Codex to transform those existing materials into a responsive Next.js application, integrate GPT-5.6 through the OpenAI Responses API, create a structured evaluation schema, add deterministic scoring and professional-risk safeguards, and test and calibrate the completed prototype.

Challenges I ran into My first challenge was entering Build Week without software-development experience and learning how to direct Codex clearly enough to build a complex working system. Other challenges included translating professional judgment into measurable criteria, preventing unsupported or steering language, calibrating GPT-5.6 responses, resolving API authentication and timeout problems, and creating visuals that supported the learning experience without distracting from it.

Accomplishments that I’m proud of I created a functioning proof of concept showing that AI can help move professional education beyond content delivery and recall-based testing. NHS Decision Lab allows a learner to practice judgment, receive criterion-based coaching, and see how incomplete reasoning differs from professionally risky guidance. I am also proud that the application passed its automated tests, TypeScript validation, production build, responsive reviews, and live GPT-5.6 calibration tests.

What I learned I learned that professional expertise can be transformed into a working interactive learning system much faster and more thoroughly when a subject-matter expert works effectively with tools such as Codex and GPT-5.6. I also learned that AI evaluation requires more than a good prompt; it needs a clear rubric, structured outputs, deterministic safeguards, testing, calibration, and a governed knowledge base.

What’s next for NHS Decision Lab This proof of concept will play an important role in a larger professional-education initiative for the real estate industry. The next steps would include additional client scenarios, a curated and version-controlled professional knowledge base, more deterministic calibration near proficiency boundaries, learner progress records, instructor reporting, and testing with real estate professionals and education organizations.

Built With

• codex
• gpt-5.6