Engineering

Fine-tuning an open-source model to match frontier quality is the easy part but serving it cost-effectively is the real challenge.

How Hebbia measures agent quality at scale with a hybrid evaluation methodology.

Hebbia researchers leveraged classic signal processing techniques to build a text detection model smaller than most of the images it classifies.

We built a statistically rigorous, consensus-based framework for evaluating LLM outputs and used it to benchmark today’s leading models on the tasks that matter most to finance professionals.

We built a multi-agent system that goes beyond public web search to synthesize insights for any data source, including proprietary data sources.

At the end of last year, we returned to the drawing board and redesigned Matrix Agent.

We built a distributed LLM request scheduler that intelligently routes billions of tokens per day across multiple providers so high-priority work always gets through, even under rate limits.

After pioneering semantic search and RAG, we found both fell short on the hardest questions so we scrapped them and built a new information retrieval system from scratch.