Fyxer Achieves 90% 90-Day Retention with AI Assistant Trained on 500,000 Hours of EA Data

Fyxer developed an AI assistant that drafts personalized emails by combining OpenAI models with over 500,000 hours of human Executive Assistant (EA) workflow data. Since launch, the product has recorded a 90% 90-day user retention rate, with 53% of AI-generated drafts accepted without modification, demonstrating high reliability.

Architecture: Specialized Model Splitting

Fyxer adopted a strategy of splitting email tasks into 30–50 specialized models instead of using a single large model. This approach addresses Moravec's paradox arising from context-dependent communication.

• Reply decision model: Classifies whether a reply is needed, scheduling adjustments, information delivery, etc.
• Intent/Outcome models: Analyzes conversational intent and predicts outcomes such as meeting scheduling or request resolution.
• Memory/Retrieval models: Determines details to persist across conversations and retrieves and re-ranks relevant memories.

Training and Optimization Strategy

Fyxer utilized subtle judgment criteria accumulated from actual EA work (e.g., immediate response vs. wait, response differences by person) as training data.

• Applied Supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific model variants and control training costs.
• Initially used the OpenAI fine-tuning platform, but recently collaborated with the OpenAI managed fine-tuning team for production deployment.
• After deployment, applied Direct Preference Optimization (DPO) based on user feedback to learn the differences between original and edited versions without manual labeling.

Performance and Future Vision

Fyxer conducts A/B tests on all draft changes and releases new versions only when statistically significant improvements are confirmed. Co-founder Archie Hollingsworth stated that the plan is to evolve beyond simple draft creation into a proactive AI assistant managing overall communication and coordination tasks.