Research
Bridging the gap between research and reality.
Our research directly informs the production systems we ship. We publish, open-source, and collaborate — because great AI isn't built in isolation.
Active Research Areas
What we're working on
Computer Vision
- Real-time pose estimation & on-device deployment
- Sports biomechanics analysis from video
- Medical imaging — X-ray, MRI interpretation assistance
- Object detection for manufacturing quality control
Language & Reasoning
- Domain-adaptive RAG for low-resource domains
- Citation-grounded decoding for hallucination-free AI
- Multi-step reasoning in agentic workflows
- Clinical NLP and medical document understanding
Efficient AI
- Model quantization — INT8/INT4 with minimal accuracy loss
- Knowledge distillation for edge deployment
- On-device inference optimization (CoreML, TFLite, ONNX)
- Neural architecture search for resource-constrained hardware
Embodied AI
- Voice AI agents for real-world task execution
- Multimodal perception for physical therapy robotics
- Sensor fusion for human movement understanding
- Feedback-loop systems for interactive coaching
Publications & Preprints
Our research output
Monocular 3D Pose Estimation for Automotive Pedestrian Intent Prediction
A lightweight monocular 3D pose estimation pipeline that predicts pedestrian crossing intent in real time on automotive-grade hardware. We introduce a temporal keypoint smoothing module that resolves jitter artefacts common in single-camera setups, achieving state-of-the-art accuracy on nuScenes-Pedestrian while fitting within the latency envelope of NVIDIA Drive Orin.
Key Results
- <12ms latency on Drive Orin
- Outperforms ViTPose-S on COCO-val
- Novel temporal keypoint smoothing
Calibration-Aware INT8 Quantization for Transformer-Based Detection Models
We propose a calibration-aware quantization strategy that preserves detection accuracy when converting transformer-based object detectors to INT8. By jointly optimizing the calibration dataset and per-layer sensitivity, our method achieves 6.8× throughput improvement on TensorRT INT8 with less than 0.8% mAP degradation on COCO.
Key Results
- 6.8× throughput on TensorRT INT8
- <0.8% mAP degradation on COCO
- Works on DETR, RT-DETR, DINO
RAG with Citation Grounding: Eliminating Hallucination in Domain-Specific LLMs
Hallucinated citations remain a critical barrier to deploying LLMs in legal and medical workflows. We introduce a citation-loss training objective and constrained decoding scheme that grounds every generated reference to a verified source document, achieving zero hallucinated citations on our legal benchmark without sacrificing fluency.
Key Results
- Zero hallucinated citations on legal benchmark
- Compatible with LLaMA 3, Mistral, GPT-4o
- Novel citation-loss training objective