Now

This is a now page. It is the public, curated slice of what I am doing. It gets updated when something changes, not on a schedule.

Working on

• Leading development of a production suite of agentic document tools for a global pharmaceutical sponsor at ZS Associates.
• Hybrid agentic pipelines that pair LLM agents with classical ML models. The agent reads the question, decides which analysis to run and calls regression, decision-tree or gradient-boosted models as tools. The models produce the numbers, so every estimate is reproducible and auditable in a way an LLM alone cannot guarantee.
• Agent harnesses that hold up in practice for autonomous software development and research: durable work queues and checkpoints, guards on what each agent may change, and independent review, so multi-agent work stays on track across long sessions.
• End-to-end ML platform architecture on AWS: ingest, versioned storage, reproducible training, model registry, serving and monitoring.

Researching

• Reinforcement learning as pretraining: using reward-driven learning as the pretraining objective itself, instead of only as a fine-tuning step after next-token prediction.

Learning

• How to work with big data efficiently and connect (ingest, store, version) data.