Keegan Choudhury
Book a chat

Product manager · Washington, DC

I'm Keegan. I build tools for working with AI agents.

I'm a product manager at Capital One. Before that I spent two years at Microsoft working on Excel and AI agents. On my own time I build Array, a desktop app where the AI agents you already use work on one project together, and a few smaller apps around it.

At Microsoft I co-wrote a KDD 2025 paper on repairing Excel formulas ↗ and co-invented a published patent application ↗.

See the projects ↓ See the evidence → Résumé (PDF)
An illustration, not a screen recording: Array on a sample project. A coordinator splits the work, Codex runs the tests, Claude drafts a brief, Gemini reviews it, and you reply. Watch the recorded demo ↓

Projects

What I'm building

Built on my own time through AI-assisted development. All three are in beta.

Array

Beta

A desktop app that keeps your AI agents, projects, documents and conversations in one workspace, with shared notes that survive when a conversation resets.

Long projects outlast a single AI conversation. Array keeps each project's decisions and unfinished work in a shared notes library on your computer. For Claude and Codex, it also saves the conversation before the context resets, so the next agent picks up where the last one stopped.

  • Works with the agents you already use. A coordinator takes the task and brings in another agent when a bounded piece needs one. I run Claude, Codex, Gemini, Grok and Muse; add others by the command that starts them.
  • Ask the team: compare a short proposal from each agent side by side before you decide.
  • A preview pane beside the conversation for the documents, pages and diffs the agents make.
  • Closing the window leaves the agents' sessions running. Scheduled tasks run while Array is open, and questions can be sent over from the Voyager browser.

I use Array to build my other projects, including Voyager. It was called Assembly until September 2026, and the recording below still shows that name.

Built with ElectronReactTypeScriptMCP

How Array works: your agents, whichever you use, report to a coordinator, which saves decisions to the project notes and brings finished work to you. Claude Codex Gemini Grok +any agent Coordinatorsplits and checks Youreview once Project notessaved before resets
Array · recorded demoSample data
Recorded September 23, 2026, when the app was still called Assembly, so the window shows that name and the interface has changed since. The projects, agents and files are fictional, and no accounts or models are connected. Read the walkthrough · Watch MP4

Voyager

Beta

A Chromium-based browser where AI agents work in the tabs you pick, and nothing on a website changes until you allow it.

  • Agents that research with cited sources, fill in forms, build comparison tables, watch a page for changes, learn a workflow you show them, or look into a broken page.
  • Every click, edit or page change shows you exactly what it will do, and each approval works once.
  • Type or click in the page and the agent pauses. Stop takes over at any time.

The newest build gives each project its own space in the side rail and three ways to see your tabs: as pages, as a wall of pictures, or as a board that sorts agent work by what it needs from you. Voyager Sync keeps your open tabs and bookmarks on each computer you add, locked before they leave. The recording shows that development build.

Illustration: an agent fills in a form, stops before Save, and waits for you to allow it once.

Voyager · development buildSample sites
A walkthrough of Voyager's development build, recorded from the real app on September 23, 2026 with fictional sample sites. The watcher and the taught workflow run on the computer, with no AI model. Read the walkthrough · Watch MP4

Dreamcatcher

Beta

Gives AI answers a paper trail: the source passage behind each claim, the reviews recorded for the answer, and a way to challenge the result. An open challenge holds back the answer's evidence score until someone reviews it.

Real-world accuracy hasn't been measured yet. Watch the recording

The illustration uses sample data.

Code

Code you can read

Source-available: anyone can read and use it; nobody may offer it in a competing commercial product.

2026Source-available · FSL-1.1-MIT

compaction-guard

A Claude Code hook for the moment an agent compacts its conversation. It saves an exact copy first, checks the copy by hash, and writes a receipt. If any step fails, the compaction is blocked and nothing is lost. Tested live in Claude Code on September 23, 2026. The story behind it

Illustration of the steps. If one fails, the compaction is blocked instead.

View the code ↗

Research

Papers and patents

From my time at Microsoft.

2025Paper · KDD 2025

Evaluating formula repair (FoRepBench)

A benchmark of 618 examples across five Excel error types, used to compare four language models. Each fix is checked two ways: does it run, and does a judge model agree it's right. I worked on the scenarios, the evaluation criteria and the writing.

Illustration of the method, not results from the paper.

Read the paper ↗
2025Patent application

Guided conversations

An agent gathers information by talking with a person, follows rules along the way, and fills in a structured document. I'm a co-inventor on the published application, US20250378320A1.

Illustration of the idea, not the patent's own example.

Read the application ↗

Work history

Where I've worked

Full résumé (PDF) ↗

Get in touch

I'm in the Washington, DC area and open to roles in the SF Bay Area, Seattle and New York. Email me, or book a time on my calendar.

Book a coffee chat Email LinkedIn

Outside work I read history and science fiction, and research my family's genealogy.