Nitesh Methani

Nitesh Methani

Hey, I'm Nitesh.

Intro:

I'm Nitesh Methani. I build the systems that make AI agents useful after the demo ends: runtimes, evals, orchestration, observability, and inference.

My path runs from machine-learning research at IIT Madras, to applied work at Amazon, to building dependable AI systems at Sarvam AI. This is my public notebook, where I think in public about the systems I'm building and the questions they leave me with.

Say hello for thoughtful technical collaborations and good conversations.

  1. Serving Kimi K3 in Production: What It Takes to Run a Trillion-Parameter Model

    A systems series on serving Kimi K3: from matrix multiplication and KV caches to GPU clusters, MoE routing, and agent economics.

  2. My notes from Baseten's Guide to Evaluating Open Models for Coding Agents

    A poor harness can make an expensive model look bad—and a cheap one look costly. Notes on measuring coding agents by cache-adjusted cost per successful task.

  3. Saved Quotes

    A growing collection of quotations worth remembering, with a note on why each one stayed.

  4. Building a managed agent platform

    A living, collaborative map of the model behaviors and platform systems that make enterprise agents dependable.

  5. Separating signal from noise in financial news

    A framework for separating consequential information from market chatter, and for knowing when a headline should actually change your view.

  6. The pessimist VC's 15-point kill list

    A red-team checklist for startups: the questions that expose fragile ideas before a compelling story takes hold.

  7. Tayari

    A GATE CS preparation app built around focused practice and recall, now used by more than 10,000 students.

  8. Learning geospatially aware place embeddings via weak-supervision

    A SIGSPATIAL paper on combining delivery history, address text, and geography to resolve ambiguous real-world locations.

  9. A Systematic Evaluation of Object Detection Networks for Scientific Plots

    An AAAI study of how object detectors read scientific plots, and why charts demand more precise localization than natural images.

  10. PlotQA: Reasoning over scientific plots

    A WACV benchmark that made visual reasoning over scientific plots a rigorous, measurable problem for machine learning systems.