Fyxer has built an AI executive assistant by integrating OpenAI models with a large corpus of annotated workflows and a modular model architecture. The company trained on more than 500,000 hours of human executive assistant work and divided email tasks across dozens of specialized models to generate context-aware drafts.
Architecture and training
Rather than treating email as a single text-generation task, Fyxer splits the workflow into 30–50 narrow models. Separate components classify whether a message requires a reply, identify intent, predict likely outcomes (for example, scheduling or resolving a request), and retrieve relevant memories for a thread. Retrieval models compare incoming messages to stored interactions and surface the details most relevant to the user and conversation.
Fyxer used supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific variants while managing training cost. The team tested models on internal validation sets for drafting, classification, and prioritization, weighing accuracy against response time and cost. Fyxer also worked with OpenAI’s managed fine-tuning team to deploy a new checkpoint into production. “We chose OpenAI because they have the best models, and they’ve given us real access and a close working relationship,” said Archie Hollingsworth, Co-founder.
Feedback loop and performance
After deployment, Fyxer converts user edits into training data using Direct Preference Optimization (DPO). The system treats corrected drafts as preferred outputs and runs A/B tests on updates, only shipping changes that produce statistically significant improvements. According to the company, 53% of AI-generated drafts are accepted as written.
The startup reports rapid commercial growth: revenue rose from $1 million to $32 million in annual recurring revenue in 2025. Retention is highlighted as a core metric—over 90% of users remain customers at the 90-day mark, and many use the service daily, Archie Hollingsworth said.
Joey Dwonczyk, AI/ML Product Engineer, said OpenAI’s support helped transfer customer-specific learnings into model behavior. Fyxer is now extending its assistant to better model relationships, preferences, and ongoing work threads as it moves beyond drafting toward broader coordination tasks.
Original source: OpenAI News