Fyxer’s AI executive assistant is the result of a multi-year engineering and product effort to replicate how skilled human assistants manage inboxes, meetings and ongoing threads. Rather than treating email as a single text-generation task, Fyxer decomposes the assistant role into a set of interlocking prediction jobs and stitches together OpenAI’s frontier models with dozens of narrow submodels, extensive fine-tuning on real assistant workflows and a live feedback loop driven by user edits.
At the core of Fyxer’s approach is modularity. The platform partitions the assistant’s responsibilities across roughly 30–50 specialized models, each focused on a specific part of the email workflow. OpenAI models handle central steps — understanding incoming messages, retrieving relevant context from memory, re-ranking retrieved items, and producing ranked candidate drafts — while narrower classifiers and predictors decide whether a reply is required, whether scheduling action is needed, or what the likely trajectory of a conversation will be.
This decomposition reflects Fyxer’s conviction that effective email automation requires more than one generator. A reply-decision classifier first determines whether a message is informational, needs a reply, or calls for scheduling. If a reply is indicated, downstream models infer sender intent, predict the conversation’s likely direction (scheduling, resolution, or continued dialog) and surface the most relevant memories. Memory and retrieval models are central to this architecture, helping the system decide which details should persist across exchanges and which items are ephemeral to the current thread.
Fyxer’s model strategy is grounded in real-world assistant behavior. Before launching the AI product the company operated a human-run executive assistant service; that operation produced a dataset of over 500,000 hours of annotated executive workflows. The dataset captures practical judgments assistants make daily: when to respond immediately, when to wait, which prior exchanges matter, and how phrasing varies depending on the recipient and context. These annotations provide the basis for supervised fine-tuning and targeted model adaptation.
To convert human workflows into reliable model behavior, Fyxer uses a blend of fine-tuning techniques and cost-aware adaptations. The team applies supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific model variants while managing training cost. Early work used OpenAI’s fine-tuning tools for high-accuracy tasks, and Fyxer later collaborated with OpenAI’s managed fine-tuning team to deploy new checkpoints to production. Models are validated on Fyxer’s task-specific test sets for drafting, classification and prioritization, with evaluation balancing accuracy, latency and cost trade-offs.
User behavior drives continual improvement once models are live. When users edit a generated draft before sending, Fyxer records the original draft paired with the user-edited final as a preference signal. The company trains on these comparisons using Direct Preference Optimization (DPO) rather than relying solely on manual labels. Candidate improvements are A/B tested and promoted to production only after achieving statistically significant gains; the product’s user volume means some experiments reach significance quickly, sometimes within a day.
Those operational choices appear to deliver real-world alignment: Fyxer reports that 53% of AI-generated drafts are accepted as-is, indicating the system often matches user intent and tone without edits. The company emphasizes retention as a core metric of trust and usefulness. In 2025 Fyxer grew from $1 million to $32 million in annual recurring revenue, and more than 90% of users remain paying customers at the 90-day mark, using the service daily.
Looking forward, Fyxer plans to extend the assistant beyond drafting toward broader coordination tasks by deepening its model of relationships, preferences and work threads. The stated vision is to allow customers to entrust the assistant with increasingly large portions of their communication workload.
Fyxer’s technical choices — modular submodels, task-specific fine-tuning on half a million hours of annotated workflows, and preference optimization drawn from everyday user edits — sketch a practical path for building trusted, contextual AI tools for professional communication. By aligning frequent model updates to how experienced human assistants actually operate, and by validating every change through user-driven A/B tests, Fyxer demonstrates a replicable strategy for deploying assistants that people both accept and continue to use.
In sum, Fyxer’s blend of modular architecture, real-world training data and a continuous, feedback-driven training loop shows how an AI executive assistant can achieve high acceptance and retention without sacrificing the nuanced judgments that make human assistants valuable.
Source: Read the original source
Leave a Reply