Feed

12 materials

Sort by

Support Architecture for Millions of Conversations: Separating Retrieval, Actions, and Generation

Using a housing booking service as an example, the authors examine the shift from a single model to a constrained orchestrator of typed operations and a generator grounded in verified context; they assess the architecture’s effects separately from accompanying changes. In replayed conversations, booking selection became more accurate, errors in structured actions disappeared, and escalations declined. The optimization also reduced latency and model serving costs, while the volume of conversations handed off to human agents in production remained largely unchanged.

Read

Playco: GPT-6 Astra helped cut manual rework of game prototypes in half

Three themed versions were made from a single graybox layout, and most of them worked on the first try. Compared to the previous model, the number of manual fixes dropped by 50%, and spatial logic, UI responsiveness in Unity, and game feel also improved.

Read

LLM coordinator for energy: one query activates an entire ensemble of specialized AI models

The new work proposes the ADN-Agent architecture, in which a large language model takes on the role of a dispatcher for a set of specialized models used in managing distribution networks. The system itself determines the meaning of a request, breaks it down into stages, selects the necessary tools, and connects them through a common interface, while automated fine-tuning of compact LLMs is used to simplify text subtasks.

Read

ChannelFlow-Tools: a customizable pipeline for generating 3D flow-around datasets for machine learning tasks

The authors presented an open pipeline that automatically generates obstacle geometry, performs hydrodynamic modeling, and converts the results into tensors for training neural networks. The system was validated for reproducibility and data quality, and the resulting dataset enabled the successful training of several surrogate models.

Read

Personal itinerary: how ChatPlanner uses LLM to take passenger preferences into account

The new ChatPlanner framework combines fine-tuned language models with RAG to recognize conversational preferences and incorporate them into route calculation. In experiments, the system found more suitable alternatives than standard planners while remaining computationally practical.

Read

How to train language segmentation on ultrasound when there is almost no labeled data

Researchers proposed a method for adapting a model to a new set of ultrasound images without access to the original data: it combines pseudo-label refinement, contour quality control, and synthetic image generation in the style of the target domain. The approach noticeably improves segmentation accuracy even when starting with just five labeled frames.

Read

AdaLens: a clear storyline to help control and redirect long-running agent analyses

As LLM agents in data science become more autonomous, standard interfaces fail to provide a clear picture of the system's reasoning or the ability to intervene in time. AdaLens visualizes the analytical process as a storyline: plans, progress, and intermediate results are visible, and the analyst can redirect the course of work.

Read

FlightLLM: How semantic language models help explain the causes of aviation incidents

FlightLLM's novel approach combines statistical models and large language models to turn flight data into clear explanations of pilot actions. Tests on real Airbus A320 data showed that the method delivers accurate classifications and coherent descriptions of the causes of hard landings.

Read

How Virgin Atlantic connects fragmented customer signals with ChatGPT Work

The airline implemented OpenAI's corporate tool to analyze passenger data and make decisions faster. This enables teams to build a complete picture from various touchpoints and accelerate the launch of new products.

Read