Airflow in Action: Customizing LLMs at Laurel. Speeds Rollout and Saves $500k / Year in Inference Through GenAI Ops

Timekeeping is a tedious but essential process for legal and accounting firms, where precision directly impacts billing accuracy. At the Airflow Summit, Moulay Zaidane Draidia, founding member of Laurel’s AI team, shared how the company leverages Apache Airflow to orchestrate and optimize GenAI models for automating timesheet creation in his session Customizing LLMs: Leveraging Technology to Tailor GenAI using Airflow.

AI-Driven Workflows for High-Stakes Billing

Laurel, a Series B startup with $55 million raised, is transforming how professionals track billable hours. Manually logging work activities is time-consuming and error-prone, especially in industries where lawyers and accountants bill in six-minute increments. Some professionals manage over 50 projects daily, making accurate record-keeping a challenge.

Figure 1: Laurel’s GenAI models accurately calculate billable hours by collecting, grouping, classifying, and summarizing employee activities from their digital footprint. Image source.

Laurel’s solution automates this process by capturing digital work activities—emails, meetings, document edits, and more—through its assistant. This data is then processed by AI models that categorize tasks, assign billing codes, and generate descriptions. However, scaling this AI-driven workflow presented operational challenges:

• Privacy constraints: Clients required strict data isolation, preventing model training across firms.
• Frequent retraining needs: Legal and accounting workflows change daily, causing significant data drift.
• Varied user behaviors: Some users log time continuously, others only at the end of the month.
• Operational reliability: The system is mission-critical—billing accuracy directly impacts revenue.

Why Airflow for GenAI Operations?

Before adopting Apache Airflow®, model management at Laurel was manual and slow. Each customer required individual model retraining and deployment, consuming valuable engineering time. Backfilling historical data or running simulations was cumbersome.

By integrating Airflow, Laurel standardized its ML pipeline into modular DAGs that automate:

• Daily model retraining to combat data drift.
• Automated model deployment based on performance evaluations.
• Cost-aware inference scheduling to balance accuracy and efficiency.

This shift streamlined operations, allowing fast iteration, safe rollouts, and improved overall system performance.

Figure 2: Airflow orchestrates multiple analytical, large language and small language models to power Laurel’s workflows. Image source.

Personalization at Scale with Airflow

Laurel’s AI strategy hinges on continuous learning from user interactions. Timekeepers review AI-generated entries, edit them, and submit final versions—creating a feedback loop that enhances personalization. However, delivering user-specific models in real time requires careful orchestration.

Airflow’s DAGs help Laurel:

• Optimize retraining frequency: DAG parameters function like hyperparameters, controlling how often models update.
• Scale efficiently: Instead of one-size-fits-all models, Laurel fine-tunes lightweight, user-specific Small Language Models—reducing training costs while maintaining accuracy.
• Cache and provision models dynamically: By predicting usage patterns, models are loaded into memory only when needed.

Unlocking Major Inference Savings with Airflow

Running LLMs at scale can be prohibitively expensive. Laurel tackled this by adapting inference schedules to user behavior:

• Fast, cheap models handle real-time predictions.
• More powerful LLMs run periodically in Airflow DAGs, summarizing daily activities just before users review their timesheets.
• RAG-based description generation is scheduled to align with billing cycles, reducing unnecessary API calls.

By shifting inference workloads to Airflow, Laurel reduced LLM costs by over $40,000 per month without sacrificing accuracy.

The Airflow and Astronomer Advantage in GenAI Operations

In addition to the Airflow Summit session Laurel also discussed how GenAI Ops had evolved at the company since the addition of Astro, the fully managed Airflow service.

“After adopting Astro, the Laurel data team now had a powerful tool to act as the center of their modern data stack, eliminating manual workloads and enabling them to start building new pipelines to better serve their business. As they began building these new pipelines and onboarded more and more customers, Astro’s auto-scaling capabilities made sure that their Airflow environments were able to handle the increased load efficiently without compromising on performance.”

You can read more in the Laurel and Astronomer case study.

Airflow and Astronomer has become the backbone of Laurel’s GenAI workflows, delivering:

• Modular, reusable pipelines for rapid iteration and experimentation.
• Safe model rollouts with controlled evaluations and deployments.
• Reducing costs with adaptive inference scheduling without degrading user experience.
• Scalability across diverse models, from traditional classifiers to cutting-edge LLMs.

Learn more about LLM operations and orchestration with Airflow and Astro from the solutions page, along with practical steps to get started.