HealthBench Evals: a desktop app built with spec-driven development
Rebuilding my HealthBench tooling as a cross-platform Flutter app, to test spec-driven development with an AI agent against a known objective.
Software engineer
carlos-dev-research
I build tools, systems and small experiments, and write about how they work.
Top projects
A lightweight harness for HealthBench evaluations
Grading language models on realistic health conversations with physician-written rubrics, through any OpenAI-compatible API.
Detecting packages with RF-DETR
Fine-tuning a transformer-based object detector on a public package dataset, evaluating it against a reference model, and shipping it as a Gradio app.
Recent posts
All posts →-
HealthBench Evals: a desktop app built with spec-driven development
Rebuilding my HealthBench tooling as a cross-platform Flutter app, to test spec-driven development with an AI agent against a known objective.
-
A lightweight harness for HealthBench evaluations
Grading language models on realistic health conversations with physician-written rubrics, through any OpenAI-compatible API.
-
Detecting packages with RF-DETR
Fine-tuning a transformer-based object detector on a public package dataset, evaluating it against a reference model, and shipping it as a Gradio app.
-
Counting items in warehouse bins on AWS SageMaker
Testing four model heads on the Amazon Bin Image Dataset counting task, and what comparing against trivial baselines revealed about the results.