jimmyjames177414
James Robert Miller
AI architect at Accenture, building its internal AI assistant. On my own time I build open-source tools that measure what AI coding agents actually do.
01 About
I'm James Robert Miller. Online I go by jimmyjames177414, on GitHub and everywhere else.
I started building software at Purdue, as lead developer on a disaster-response system funded by Homeland Security and Department of Defense grants and as an application developer at No Magic, then spent three years on Saggezza's UX team before joining Accenture in 2018.
Today I'm an AI architect on Amethyst, Accenture's internal AI assistant, and one of the developers who originally built it at a 2019 hackathon. I own its orchestration layer across Amazon Bedrock, Azure OpenAI and Google Vertex AI. Lately I spend as much time on how teams build with AI as on what they build: context is an asset, and an AI is only as useful as what your team has written down.
On my own time I build tools that measure what AI coding agents actually do: what context they needed, what they changed, what they broke, and what they quietly got wrong.
Each of the eleven research tools ships a written prior-art search, including a list of claims the project is not allowed to make. Outside all that: Arduino, Raspberry Pi, and taking things apart to understand them.
$ whoami James Robert Miller (jimmyjames177414) $ cat focus.txt tools that measure what AI coding agents actually do $ ls projects/ rollcall sincelast stopless mincontext errorbars knowwhen validwhile toolsweep worker-king ... $ echo $EMAIL jimmyjames177414@gmail.com
02 Projects
- Watching agents work
rollcall
Which worktree each running AI coding agent owns, and which work nobody owns. Read the write-up.
- Watching agents work
sincelast
Tells a coding agent what changed underneath it since it last read a file.
- Measuring context
stopless
Which words a prompt actually needs, measured by real ablation.
- Measuring context
stopbench
A contributable benchmark for how much language an LLM needs.
- Measuring context
mincontext
Delta debugging for AI failures: shrink a failing context to the smallest reproducer.
- Measuring context
toolsweep
Which tool-schema decision is costing accuracy, by factorial ablation.
- Honest evaluation
errorbars
Power analysis and paired confidence intervals for LLM evals. Read the write-up.
- Honest evaluation
inconclusive
Did the property fail, or did the run just end? Three-valued verdicts over agent traces.
- Honest evaluation
knowwhen
A benchmark for whether an agent knows when to ask instead of act. Read the write-up.
- Accountability
assumptionledger
Makes an agent state its assumptions in a form you can check before it acts.
- Accountability
validwhile
Agent memory where every fact carries a check for when it stops being true.
- Other
worker-king
Always-on AI desktop assistant for Windows that delegates work to Claude Code.
- Other
claude-roundtable
Two Claude personas argue a topic adversarially, streamed live in the terminal.
- Other
claude-memory-atlas
Renders Claude Code memory files as an offline, interactive knowledge graph.
03 Experience
From disaster-response software at Purdue to enterprise AI at Accenture, plus what I build in the open. More on LinkedIn.
- 2009-2013
Purdue University Calumet
Bachelor's degree, Computer Science through CIT&G
Strong enough in databases and full-stack development to take on two technical roles while still a student.
- 2010-2016
Purdue University
Full-stack .NET developer, DMCS
Built the Disaster Management Communication System alongside its founder, Professor Barbara Nicolai: my first full-stack production application, in ASP.NET, C# and SQL Server, on a program funded by Homeland Security and Department of Defense grants. It was built to help first responders and the wider community coordinate and close out an incident quickly. Covered by Purdue news and in the January/February 2015 issue of Building Indiana.
- 2013
Open source
Joined GitHub as jimmyjames177414
Same handle ever since.
- 2013-2014
No Magic, Inc. · Allen, Texas
Application developer, No Magic
Built a drag-and-drop dashboard system for MagicDraw, No Magic's UML modeling software. The demo drew the CEO and earned a full-time offer, which I turned down because I had just been accepted into a master's program. My first commercial Agile team.
- 2015
Saggezza
Won a place on the UX team
In onboarding, six new starters rebuilt a website from a single screenshot. Mine was the only pixel-perfect one, which earned me a spot on the design-led UX team.
- 2015-2017
Saggezza · Chicago
Senior full-stack UX developer
Delivered projects end to end, on teams and solo, for clients including Cigna, MasterCard, P.F. Chang's, RR Donnelley, LSC Communications, VSP Direct Insurance and Calero. Built Saggezza's corporate and UX websites and the OneSaggezza mobile app (Ionic and Angular, Android and iOS).
- 2015-2017
Saggezza
Two weekends for accessibility
A colleague's project, not mine, failed the WCAG AA standard its contract required, because the team had been told to scan with a different tool. I volunteered two weekends and fixed it.
- 2015-2017
Saggezza
Printing technology for RR Donnelley
Extended the XBITS XML format so printing equipment from acquired companies could talk to RR Donnelley's systems. Also built an iPhone app that read Morse code from a blinking light through the camera, to pull data off RR Donnelley's printed electronics: a sticker that blinked its data when you pressed a button on it.
- Feb 2018
Accenture · Chicago
Joined Accenture as a senior software developer
First project: KX Documents slide-level search, so people search inside PowerPoint decks and land on the most relevant slide, not just the best document.
- 2019
Accenture
First hackathon win: the chatbot that became Amethyst
One of the original creators of the AskAccenture chatbot. It began as a hackathon our team won and reached production as an MVP within two months, then became OmniBot and later Amethyst, Accenture's internal AI assistant. I built the employee lookup and PTO capabilities. Same year: merged a fix into Microsoft BotFramework-WebChat (PR #2331).
- 2020-2023
Accenture
Special Projects and Custom Engineering
Built the people-lookup app on the web and in Microsoft Teams for employee discovery, a rapid single-page-app deployment platform on AWS S3 that cut delivery from months to weeks, and moved the internal survey platform from Windows Server and IIS to Docker, Alpine Linux and Nginx.
- Sep 2023
Accenture
Back on Amethyst, the last original member still on it
Worked on its move from intent-based Azure LUIS to a GPT-based 2.0: backend, plugins and enterprise API integration with Semantic Kernel, on a platform of C# Azure Functions with Angular and Microsoft Teams front ends.
- 2023-2024
Hackathon
Second hackathon win
Two-time hackathon winner.
- Dec 2024
Accenture
Promoted to architect
First architect assignment: repeatable performance testing with JMeter and Azure Load Testing in containers, provisioned with Terraform and wired into pipelines.
- 2025
Accenture
Technical architect for a global team
The US-based architect working with engineers in Manila and testers in Argentina. Worked across agent registration and orchestration, access-controlled agent selection, the .NET 8 move for Azure Functions, and integrations including Microsoft Graph, M365 Copilot, Teams SSO and Adobe Firefly.
- Jul-Sep 2025
Accenture
Isolated environments for agent builders
Designed and built two separate environments, each with its own databases and AI services, so teams building agents on Accenture's agent platform can experiment without touching internal dev and staging.
- 2025-2026
Accenture
Voice, safety controls, testing and enablement
Shipped text-to-speech and speech-to-text, then real-time voice on AWS Nova Sonic: a spoken conversation you can interrupt mid-sentence. Added security and compliance validation at the user and response level, rebuilt the business testing tool on worker processes, and automated conversion of about 1,000 Markdown files to Word. Ran hands-on enablement sessions across teams, and reviewed and signed off the team's code changes.
- 2026
Accenture
Multi-cloud orchestration and agent testing
Own the orchestration layer across Amazon Bedrock, Azure OpenAI and Google Vertex AI, plus the conversation, workspace and memory services and the retrieval and ingestion pipeline. Built and shipped automated testing for AI agents: repeatable runs with stored results so one release can be compared with the last, and overlap detection that catches two agents covering the same subject before users do.
- Early 2026
Accenture
Showcased Amethyst in Chicago
Represented the platform at an Accenture Executives Club booth in Chicago.
- Jul 2026
Open source
worker-king
An always-on desktop assistant that hands real work to Claude Code.
- Aug 2026
Accenture
How engineering works when AI writes code
On a leadership-sponsored pilot of a GenAI-native engineering team, owned AI tooling and shared context and wrote the ways-of-working standard the team builds by. That became a shared team memory: decisions, meetings and requirements captured as they happen, linked and versioned, answerable in plain language from Microsoft Teams.
- Sep 2026
Open source
Eleven tools for measuring AI agents
Four research tracks, each tool shipped with a written prior-art search and the claims it is not allowed to make.
04 Contact
Let's talk.
Email jimmyjames177414@gmail.com, find me as jimmyjames177414 on GitHub, or connect on LinkedIn.