Building a Career Co-pilot for Disadvantaged Students: How Zero Gravity Bridges Knowing and Doing

This podcast episode features Elliot Little and Dan St. Paul of Zero Gravity, a UK platform helping state-school students reach elite careers through mentoring and learning pathways. They describe building an AI career co-pilot that orchestrates a student's next steps rather than automating their work, aiming to close the gap between knowing what to do and actually doing it. The team recounts abandoning ambitious ideas like AI mentors and avatars for a simpler, more effective product, and shares technical lessons on context management across multi-month journeys, model selection, and safeguarding for 16-year-old users. The discussion matters for anyone building AI products that augment rather than replace human relationships, especially in sensitive domains like education.

01Key takeaways

• Start from the real user problem and be willing to scale back ambitious visions to something that demonstrably works.
• Make AI personalization visible to users when it is the core value, since hiding it can erode trust and engagement.
• Use guided prompts rather than blank text inputs to lower the barrier for users unsure what to ask.
• Manage context deliberately over long user journeys by pruning stale tool results and summarizing history.
• Use application logic to control tool availability and context freshness before reaching for complex retrieval architectures.
• Treat safeguarding as a core product requirement when serving minors, combining automated moderation with external checks.

02Key sections

Zero Gravity's mission

The team explains how Zero Gravity addresses unequal access to elite careers for disadvantaged UK students, using mentoring, community and learning pathways. The framing centers on a 'knowing-doing gap' that motivates the product.

From grand vision to simpler prototype

Early ideas included AI mentors and synthetic avatars, but the first prototype, a job suitability summary, did not produce the expected 'wow moment'. The team iterated toward something more practical.

Designing the interaction

The team chose text chat over voice and used guided prompts instead of empty text boxes. They also found that hiding the LLM's personalization backfired, because students needed to feel it.

Context and model management

To support multi-month journeys without exploding token counts, they removed stale tool calls, summarized history, exposed tools conditionally, and matched models to tasks. Application logic was used instead of complex RAG.

Safeguarding and evaluation

Safeguarding was treated as a first-class concern, combining moderation endpoints with external verification and a failure taxonomy built through red-team and green-team exercises. Future work focuses on long-term memory.

03From the post

“Listen to this episode on: Spotify | Apple Podcasts How do you help disadvantaged students take action on opportunities they don't even know exist? In this episode of Just Now Possible, Teresa Torres talks with Elliot Little (Product Manager) and Dan St. Paul (Software Engineer) from Zero Gravity, a UK-”

04Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.