
Blog
Research, systems, and things learned while building.
I write about model research, post-training, inference, retrieval, and the engineering details that decide whether an AI system works outside a notebook.
Writing
4 posts- Credit Assignment Is the Whole Game: A Modern Map of Reinforcement Learning from Pong to Post-TrainingA researcher-engineer guide to policy gradients, PPO, GRPO, RLOO, process rewards, and tool-use RL through the lens of credit assignment.
- When Your Multimodal RAG Team Punishes Good Retrieval: GraMRAG and Topology-Aware Policy OptimizationHow graph memory and topology-aware policy optimization assign credit across long-horizon multimodal retrieval trajectories.
- Multilingual tokenization without shortcutsDesign notes from building indicTok across 22 Indian languages and 12 scripts.
- Evidence-first RAG for high-stakes workflowsHow retrieval, citations, and human review fit together when plausible text is not enough.