Welcome — what I’ll be writing about

Posted May 04, 2026

This is the first post on what I hope becomes a running set of notes on the things I think about most days: trustworthy AI systems, the surprisingly subtle problem of memory and personalization in agentic workflows, and the gap between what models can do in a benchmark vs. what they actually do in production at scale.

A few topics I’d like to dig into over the coming months:

• Bidirectional alignment in practice — what changes when you stop treating alignment as a one-way constraint and start treating it as a contract between system and user.
• Memory as a research surface — why long-term memory in agents is more than a vector DB, and what we keep getting wrong about it.
• Evaluating agentic systems — how reference-free evaluation actually works (or doesn’t) once your system has tools, multi-turn state, and a moving target.
• From paper to product — the unglamorous middle layer between a NeurIPS result and a feature that ships in M365.

If any of this overlaps with what you’re working on, I’d love to hear from you — find me on LinkedIn or Twitter.

— Reshmi