AI model development for specific outcomes.

Define the task before changing the model

A specialized model needs a concrete target: the inputs it will receive, the outputs it should produce, and the conditions under which those outputs are useful. The target may involve structured analysis, a decision workflow, or operation within device and latency constraints.

We start by examining the task and the baseline model. Evaluation should show which failures matter and whether training is an appropriate response. Improvements need to be checked against examples that were not used to fit or tune the model.

Training and fine-tuning

Model development involves choices about training data, supervision, output format, model size, and deployment constraints. A larger model or a longer training run does not automatically address the behavior a product needs.

We work with training and fine-tuning for specific outcomes. Useful development records connect a model version with its data and training configuration, the evaluation set, and the differences observed against the baseline. This makes it easier to investigate a regression and decide which version to retain.

NommAI: calorie counting and food analysis

NommAI is an example of our specialized model work. It uses food photos to estimate calories and macronutrients and supports reviewing the result before it is added to a food diary. The product is designed around on-device AI models.

Grape estimates whole-meal totals. Plum adds food-by-food analysis on supported devices. These are different product requirements, with implications for model behavior, output structure, memory, and inference performance.

A photo-based nutrition estimate is not a direct measurement of the food’s nutritional content. Portion uncertainty and ingredients that are not visible in the photo affect interpretation. The application needs to communicate that uncertainty and let the user review and correct an estimate.

Evaluations, benchmarks, and decision modeling

We use evaluations to determine whether training changes the behavior that matters for a task. This includes examining both aggregate results and the individual examples on which a model fails. A single headline score can hide a problem in a relevant category.

For decision-oriented tasks, evaluation should describe the decision being modeled, the available information, and the criteria used to judge an output. It is important to distinguish an association in the data from a conclusion the model can support on a new case.

Benchmarks provide a shared comparison under their own conditions. Product evaluations add the actual constraints: the input distribution, output format, human review requirements, and operating environment. Both should be interpreted with their scope and limitations stated.

Connect development with research and adoption

Our AI safety research examines model behavior and agent actions. Those findings can inform what to test during model development and which observations to retain when a model is put into use.

Our enterprise advisory work addresses a related decision: whether an organization should use an existing model, develop a specialized one, or change the workflow around it. Training is part of an implementation only when it addresses a defined requirement and can be evaluated against it.