The false positive problem: how LLM judges mislead table OCR benchmarks
We built a benchmark where ground truth is known by construction, then used it to show that frontier LLM judges hallucinate table OCR errors at non-trivial rates.
Jean Ghislain Billa · Recent posts
How We Cut Parsing Latency 3x by Building Our Own Orchestration System
Bursty, latency-sensitive traffic exposed the limits of per-tenant overflow pools. So we vendored a small task queue and built our own scheduling and traffic-management layer on top, keeping shared capacity warm, fair, and policy-aware while cutting parsing latency about 3x.
Thomas Boser Fixing the Real Bottleneck in Our CV Serving Stack
Customer bursts exposed a weakness in our CV serving stack that average traffic mostly hid: the bottleneck was not model compute, but the pipeline around it.
Karan Brar How we built our agent-native observability stack
Most AI-for-observability demos start with a chatbot over logs. We cared about a less flashy question: could an agent understand our production system well enough to help operate it?
Raunak Chowdhuri Build with us.
We're hiring engineers who want to work on document intelligence, infrastructure, and the systems that make AI useful in production.