Project
Five-lane request router
A published ModernBERT classifier that routes requests among five task paths, with its training data, evaluation and tutorial.
A personal AI assistant was routing every request through a 9B generative model just to select a processing path: chat, image, search, audio or video. A fine-tuned ModernBERT-large classifier was integrated as a TensorRT FP16 engine on an NVIDIA Jetson Orin Nano for that task. At the model-card release, the 9B model still served live traffic.
Public, rerunnable evaluation: on a frozen 60-row authored challenge, the pinned router scored 58 / 60 against 46 / 60 for a predeclared TF-IDF baseline. The tutorial includes the scoring steps and raw results. It is a teaching and regression set, authored with knowledge of the label rules, not an independent blind holdout or live-traffic sample.
Separate sealed comparison: the model card also reports 60 held-out cases, run three times for each router: 180 / 180 correct calls for ModernBERT and 175 / 180 for the 9B router. Client-side p50 / p95 latency was 42 / 45.6 ms versus 771 / 1,269 ms on different hardware. That comparison is one authored test set, not a production soak or a general model-speed benchmark.
What I’d point a reviewer at: the failure that shaped the data. An earlier candidate failed its sealed test by sending a chat request that mentioned drawing to image. The fix was 253 contrast rows written for that failure shape, without seeing the failing prompts, followed by a fresh sealed test.