Insights/Research
Research
Notes from our engineering team on how we build, train and evaluate AI systems.
RESEARCH 0211 MIN READ
72%
of errors caught by word-level labels, vs 48% for whole-segment scoring
Speech-to-text · Evaluation
How We Built a Jury of LLMs That Listens Before It Judges
A jury of three AI models from different vendors listens to the audio, reads the transcript, and flags the segments a human reviewer needs to see, pointing at the exact words it thinks are wrong.
RESEARCH 017 MIN READ
2B
parameters in Qwen2-VL-2B, trained with GRPO and LoRA
Computer vision · Reinforcement learning
Teaching a Small Vision Model to Look Twice: Inside Our Detection RL Gym
In our Detection RL Gym, a small 2-billion-parameter model (Qwen2-VL-2B) works through an image the way a human annotator does: it zooms in on small objects, tightens a loose box, deletes a wrong one, and submits only when it is ready.
OUR SERVICES
AI Optimization Services and RL Gyms
Expert QA evaluation pipelines and trainable RL environments