About
Four years putting AI systems into production, alongside the people who use them.
The day job
Forward deployed at a Navy R&D lab.
I am a Computer Scientist at the Naval Undersea Warfare Center in Newport, Rhode Island, where I have worked since 2022. I run technical discovery with program-office and senior technical leadership on the organization’s AI roadmap: gathering requirements, reading the usage data coming back from what is already in service, and telling them which of their problems are not actually AI problems.
I brief and demonstrate AI capability to the lab’s customers, to Navy program offices, other labs and commands, and contractor teams, translating between technical and non-technical stakeholders. Who gets to use this technology is mostly a translation problem.
The work behind those conversations is ordinary engineering. I architected a shared platform component that more than one project now builds on, so the design decision was made once rather than repeated by every team that needed it. I built an LLM application end to end, Python services underneath and the TypeScript interface people actually touch on top, now in sustained use by people outside my team. I built the ingestion path into existing Navy data systems and sensor and test equipment I do not own, against interfaces fixed long before I arrived.
I set acceptance criteria before the build rather than arguing them after, then led four to six Navy engineers through the deployment of a machine-learning data-analytics application, its demonstration to government users, and a handoff of the working prototype on a repeatable CI/CD pipeline. Leaving them able to run it without me was the deliverable.
I hold a B.S. in Computer Science from the University of Rhode Island, earned in 2022.
The other thread
What I build on my own time.
Arcus Code is a hardened distribution of opencode, the open-source TypeScript coding agent. It carries a per-request model-routing layer upstream does not have, vLLM as a first-class provider so the agent can run on self-hosted open-weight models inside a customer’s network, and a rebuilt Model Context Protocol tool-calling path.
CareerAgent is a fourteen-service agent platform in Python and TypeScript, built as a reusable substrate and then specialized into a vertical product without a rewrite. Its agentic workflows run behind a four-gate pipeline, which is the part I would point at first.
Both are written up in detail on the projects page, along with the earlier work they are built on.
How I think about this
What I believe after building these systems.
- 01
Tell them which problems are not AI problems.
The skill is not reaching that judgment. It is delivering it so the person stays in the conversation.
- 02
Acceptance criteria before the build, not after.
Arguing about what done means once the work is finished is how a deployment stalls in review.
- 03
The permission engine matters more than the orchestration loop.
Deciding when a human belongs in the loop is what separates an agent people trust with real work from one they quietly stop using.
- 04
The evaluation decides whether it works.
Objective per-case verification, with no model grading its own output. I write the harness before I tune anything.
- 05
Leaving them able to run it without me is the deliverable.
A prototype that only works on the day it is demonstrated has not been delivered.
Capabilities
What I work with.
Delivery
- Technical discovery and requirements gathering
- Briefing technical and non-technical stakeholders
- Solution architecture
- Acceptance criteria and demonstration
- Handoff on repeatable CI/CD pipelines
- Work inside controlled-data regimes
Agents and LLM systems
- Agent orchestration, tool calling, subagents
- Per-request model routing on cost, latency and capability
- Retrieval-augmented generation on pgvector
- Model Context Protocol integration
- Evaluation harnesses with objective verification
- Supervised fine-tuning and reinforcement-learning paths
Engineering
- Python, TypeScript, C++, Go, SQL
- FastAPI, React, PostgreSQL, Docker
- vLLM, Amazon Bedrock, AWS
- PyTorch, supervised and unsupervised methods
- Integration against fixed external interfaces
- Hardware and software boundary debugging
Details
- Citizenship
- U.S. Citizen
- Education
- B.S. Computer Science, University of Rhode Island, 2022
- Based in
- West Warwick, RI
- Working
- Remote-first, will relocate, open to customer-site travel