Papers &
projects
Research papers, technical reports, systems work, and open-source contributions across agentic LLM post-training, inference, CUDA, reinforcement learning, and video generation. Links go to the paper, report, code, or upstream contribution whenever one is available.
Six documented research and engineering records
Model & systems
Occamy-1.0
A 35B-A3B co-work model for long-horizon, tool-using tasks
- 85.93%
- ClawEval strict Pass@3
- −38.6%
- Tokens per trajectory
- Agentic post-training
- Long-horizon tool use
- RL infrastructure
Research paper
CineFlow
Dependency-driven parallel execution for scalable, consistent video generation
- 5.5×
- Peak end-to-end speedup
- +17.3%
- VBench visual quality
- Video diffusion
- Parallel systems
- Dependency scheduling
GitHub · profile preview Open-source engineering
Open-Source Systems Engineering
Training reliability, long-context kernels, RL infrastructure, and inference serving
- 6+
- Selected upstream patches
- 128K
- SFT sequence length
- Distributed training
- Long-context kernels
- RL and inference systems
GitHub · repository preview Technical report
Dynamic Prefill Optimization
Adaptive online packing for lower-TTFT LLM serving
- −20%
- Time to first token
- p95
- Online control signal
- LLM serving
- Online packing
- Tail-latency control
GitHub · repository preview Systems project
FlashAttention-style CUDA Optimization
Memory-efficient kernels and fusion for GPT-2 inference
- 10×
- Lower HBM traffic
- 157
- FLOP per byte
- CUDA kernels
- Efficient attention
- GPU profiling
Undergraduate thesis
Reinforcement Learning for Legal Reasoning on Multi-Choice QA
Hybrid Zero-RL, distilled-CoT SFT, and GRPO for structured legal reasoning
- 57.6%
- Test accuracy
- 3-stage
- Hybrid training pipeline
- Reinforcement learning
- Legal reasoning
- Chain-of-thought distillation