Amit Arora

Agentic AI infrastructure, MCP, and agent evaluation.

Principal Solutions Architect, AI/ML at AWS

Adjunct Professor, Georgetown University

I work on the infrastructure layer of enterprise agentic AI: the gateways and registries that let agents discover and call tools under real authentication and audit, and the benchmarks that tell you which model to trust for a given job. Most of it ships in the open.

This site is where I keep notes on what I am learning. Papers I have read, courses I have built, and things I got wrong the first time.

Explainers

Long form pieces and infographics, each one a page of its own.

Jev: a model that decides instead of writingexplainer

How a typed, sub-second classification model works, where it breaks, and a script you can run against your own files. Sep 2026.

Jev posterinfographic

Two page print poster version of the same material.

Costing an AI-powered SQL filterexplainer

How the roofline model predicts the latency of an LLM-backed SQL filter from FLOPs, bytes and GPU limits, with no profiling run. Oct 2026.

Agentic AI bookmarks, 2026reading list

94 posts on agents, harnesses, cost and context, each read and summarized, searchable and filterable by tag. Oct 2026.

On AI: three ideas that are informing my thinkingexplainer

Deutsch on an infinite frontier, Kapoor and Narayanan on normal technology, Tao on one shared space, each checked against my own work. Oct 2026.

Self-improving agent harnesses: RRSI and Harness²explainer

Two Google Cloud AI Research papers where an editor model rewrites an agent's harness, not its weights, explained together. Oct 2026.

From unit tests to proofs: six ways to check codeexplainer

Example tests, property-based testing, exhaustive checks, state machines, Z3 and Lean, climbed on one SSRF guard, and why it matters when agents write code. Nov 2026.

Six ways to check code: posterinfographic

Two page print poster: the story on the front, the code and evidence on the back.

e-means: k-means in a tenth of the passesexplainer

Meta FAIR's mini-batch k-means beats Lloyd's 25-pass score in about two passes, with two formulas that replace hand tuning. Nov 2026.

Building

Open source, mostly infrastructure for agents.

mcp-gateway-registrylead

Enterprise MCP gateway and registry for agents, MCP servers, and skills. OAuth and OIDC, dynamic tool discovery, auditable tool access. 180+ enterprise deployments.

hal-harnessICLR 2026

The Holistic Agent Leaderboard, a standardized harness for evaluating AI agents, built with Princeton researchers.

FMBench

Foundation model benchmarking across hardware and serving stacks, so teams pick on measured cost and latency rather than intuition.

Coding harness benchmarks

18 models over 21 real tasks, reduced to a cost and quality Pareto frontier. Ships a router that names the cheapest model clearing the bar.

my-ai-assets

Prompt templates and reusable Agent Skills, including a Feynman technique framework for reading academic papers and a structured stock valuation workflow.

jev-samples

Runnable samples for Jev, TypeSafe AI's System One model. Send state plus typed questions, branch on the value that comes back. Includes a Go port that compiles the check into one static binary.

Learning

Notes, wikis, and courses I keep in public.

Knowledge base

An interlinked wiki on generative AI, compiled from clipped primary sources. 63 articles and counting.

Reading list

What I am reading, and what stuck.

AI learning path

A self paced six week program to a working 200 level grasp of modern AI, three weeks of it on agents.

Kubernetes, zero to capstone

From no containers to running the MCP Gateway Registry and serving an LLM locally, in 13 phases.

Blueberries in my salad

A book project.

Teaching

Georgetown University, M.S. in Data Science and Analytics.

I created and teach DSAN-6725, Applied Generative AI for AI Developers, one of the first comprehensive generative AI courses offered in academia. It is now the most popular course in the program and earned a perfect 5 out of 5 in student reviews for Spring 2026. I also teach DSAN-6000, Big Data and Cloud Computing, and am building a new graduate course on AI agents. 500+ graduate students so far.

Writing

Selected posts from thirteen on the AWS blogs.

All thirteen posts →

Talks

Available to speak on enterprise MCP architecture and agent evaluation.

  • 2026 ICLR  ·  Holistic Agent Leaderboard, co-presented with Princeton University researchers
  • 2026 ACM AgentSkills, CAIS Workshop  ·  From Observed Reasoning to Stable Skills
  • 2026 MCP Dev Summit, NYC and Amsterdam  ·  MCP Gateway and Registry
  • 2025 AWS re:Invent  ·  AIM3314, observability for agentic applications
  • 2024 AWS re:Invent  ·  NFX307, benchmarking FMs across hardware platforms, with Netflix