Fyxer AI Executive Assistant Technical Implementation

Fyxer implements a multi-model system to solve high-context communication

Fyxer has developed an AI executive assistant that manages professional communication by combining OpenAI models with a proprietary dataset of 500,000 hours of annotated executive assistant workflows. This approach addresses "Moravec’s paradox" — where tasks humans find easy, such as nuanced email communication, are difficult for AI — by learning individual user preferences and relationship contexts.

Decomposing email workflows into specialized models

Fyxer replaces the single-prompt approach to email generation with a system of 30–50 specialized models. Each model is responsible for a narrow segment of the workflow to increase precision and reliability.

The classification and intent pipeline

When a new email arrives, the system follows a specific sequence of predictions:

  1. Reply Decision: A model classifies whether the message requires a response, a scheduling action, or is simply informational.
  2. Intent Analysis: If a response is required, models analyze the intent and predict the likely outcome (e.g., scheduling a meeting or resolving a request).
  3. Context Retrieval: Retrieval models compare the new email against stored interactions to surface relevant memories and historical context for that specific person and thread.

Model roles

OpenAI models are utilized across the entire pipeline for digesting emails, re-ranking retrieved context, and the final generation of the email draft.

Training on human-expert workflows

Fyxer leverages its history as a human-powered assistant service to train its AI. The company uses a dataset of over 500,000 hours of human executive workflows to capture the subtle judgments required for professional communication, such as timing and tone.

Fine-tuning and optimization

To maintain accuracy while controlling costs, Fyxer employs several technical strategies:

  • Supervised Fine-Tuning (SFT) and LoRA: Fyxer uses Low-Rank Adaptation (LoRA) to create task-specific model variants.
  • Managed Fine-Tuning: The team collaborates with OpenAI's managed fine-tuning team to move new checkpoints into production.
  • Validation Sets: Every model is evaluated on internal validation sets focusing on drafting, classification, and prioritization, balancing accuracy against response time and cost.

Continuous improvement via user-feedback loops

Fyxer uses real-world user interactions to create a self-training loop that improves model performance over time.

Direct Preference Optimization (DPO)

When a user edits an AI-generated draft before sending, Fyxer treats the original draft and the edited version as a pair of outputs. The system uses Direct Preference Optimization (DPO) to learn the user's preference from these edits without requiring manual labeling.

A/B Testing and Performance Metrics

New model versions are only deployed after A/B testing shows a statistically significant improvement. This iterative process has led to the following results:

  • Draft Acceptance: 53% of AI-generated drafts are accepted as written.
  • Business Growth: Annual recurring revenue (ARR) grew from $1 million to $32 million in 2025.
  • Retention: Over 90% of users remain paying customers at the 90-day mark.

Future direction: Proactive coordination

Fyxer is evolving its system from a drafting tool into a proactive assistant capable of managing broader communication and coordination workloads, with the goal of reducing the need for users to manually interact with their computers for routine coordination tasks.

Sources