Ravenor Davion
I research transformers and study computer science at Stanford.
Experience
Research Intern @ NVIDIA
Scaling RL environments.
Researcher @ Stanford AI Lab | Advised by Christopher Ré and Azalia Mirhoseini
Post-training agents for intelligence-per-joule.
- Researcher @ Robust Systems Group | Advised by Subhasish Mitra Accelerating transformer inference.
Awards
YC Summer Fellow '26
$195k cash + compute grant for research.
Research
Privileged Context as Drift in On-Policy Self-Distillation
Demonstrated privileged context curation can reduce forgetting ~80% from naive base during OPSD; NeurIPS 2026.
AR-saLM: A fine-tuned AutoRater for search-augmented LMs
Fine-tuned Qwen 3.5 9B to judge context sufficiency - 90% accuracy at 50% lower token cost. Thanks to Barry Z. and Cyrus R. for mentorship.
Ultradense IC for Accelerating Transformer Decode
Proposed + simulated custom NoC that reduced inference latency in Llama3-9B 10.9x from SOTA.
Get in Touch
I like chatting about research and early-stage companies. I also invest on behalf of Contrary as a VP.
- GitHub rdavion [at] stanford [dot] edu
