Andrew Thompson is chief technology officer of Orbital, where he leads development of Orbital Copilot, an AI assistant for real-estate legal due diligence. His work tackles a difficult combination: rapidly improving language models, complex property documents, and lawyers’ need for accurate answers that remain traceable to their sources.
Thompson studied at the University of British Columbia before building software across financial technology and property technology. As vice president of engineering at Yoyo Wallet, he worked on the Caffè Nero app; he later became chief technology officer of Appear Here, a marketplace for short-term retail space, and founded an AI company. His experience with property transactions eventually led him to Orbital, where he helped shift its technology from classical machine-learning systems trained on lawyer-labeled documents toward language models capable of applying legal expertise more directly.
Orbital Copilot converts scanned deeds, leases, and other paperwork into machine-readable text, divides legal assignments into smaller tasks, and produces reports containing citations that lawyers can check against the original documents. Real-estate lawyers develop the system’s domain-specific instructions, while AI engineers design its orchestration, tools, and underlying behavior.
Technical ideas and practical contributions
- The prompt tax of model upgrades. Thompson describes prompt tax as the engineering and legal-expert effort required when newer models disrupt instructions and workflows that previously worked. His AI Engineer World’s Fair talk traces migrations from GPT-3.5 and GPT-4 through newer reasoning models, which often perform better with concise objectives than elaborate step-by-step prompts. He favors rapid experimentation, staged rollouts, and customer feedback to catch regressions without delaying useful capabilities.
- Legal engineering grounded in professional judgment. Thompson embeds experienced real-estate lawyers directly in product development, giving them responsibility for translating professional knowledge into prompts and reviewing outputs. Effective legal engineering must account for contractual nuance, transaction context, writing quality, and whether every citation supports its claim.
- Evaluation tied to actual legal work. Orbital initially relied heavily on expert review and user feedback because legal accuracy, concision, style, and citation fidelity are difficult to measure together. Its engineering team subsequently developed blind evaluations using real production workloads, extending that earlier approach with more structured testing.
- Product engineering built around stronger models. Thompson’s product-engineering principles emphasize shipping early and anticipating that models will become more capable, faster, and less expensive. He also advocates agentic coding within smaller, cross-functional teams: engineers increasingly define problems, direct AI coding systems, and apply product judgment. His account of Orbital’s engineering model and public LinkedIn post extend that idea to lawyers, whose AI-generated prototypes can reveal unmet needs while still requiring professional governance and maintenance.
