When prompting isn’t enough: How Doximity trained beyond the frontier

As OpenAI moved to phase out the supervised fine-tuning Doximity was using, it suggested prompting would be enough. Doximity’s production experience told a different story.

For Doximity Ask, answering clinicians’ medical questions meant delivering both clinical accuracy and fast responses. Using Fireworks, the team trained open models using its clinical expertise and physician feedback, then serve those models at the speed clinicians needed.

Doximity Ask went on to outperform GPT-5.6 Sol and Claude Fable 5 in NOHARM, an independent clinical safety evaluation.

Join Louis Mullie, MD, Chief AI Researcher at Doximity, and Vivek Chauhan, Training Product Lead at Fireworks AI, to hear how Doximity took ownership of its models to outperform the frontier in its domain and deliver answers at the speed clinicians need. Learn what it takes to improve model behavior where prompting falls short, measure whether training is working, and take greater control of what your product can do next.

You’ll learn how to:

• Turn your domain expertise and user feedback into training that improves model behavior
• Measure whether training improves quality on the tasks your users actually perform
• Build a competitive advantage by training models to outperform the frontier on the tasks that matter to your business

Part of the Build like a Frontier Lab series