Agentic systems
Building multi-agent systems with Google ADK and the Claude SDK, then integrating them into enterprise platforms.
Director of AI Technology · Dallas, TX
Aidan Marshall
I build agentic AI systems, the production tooling around them, and the controls that let teams use them safely.
I build and lead the systems that take agentic AI from prototype to production. My work spans high-stakes environments where useful automation still needs clear evidence, cost controls, and human judgment.
Director of AI Technology at Stealth AI Startup · Previously PwC and IBM
Enterprise AI delivery
Shipping AI into legal, tax, and compliance environments where governance, cost, and reliability shape the design.
AI-assisted development
Building workflows around Cursor, Claude, and Copilot that shorten implementation and make review easier to measure.
Experience
Where I've done the work
- Working with an exceptionally talented group of people on a new AI venture.
- More details will be shared when the time is right.
- Built multi-agent systems with Google ADK and the Claude Agent SDK on a globally distributed tax platform generating more than $5B in annual revenue.
- Led the 0-to-1 integration of Harvey AI into the core product, owning system design, data pipelines, and evaluation for specialized legal and compliance capabilities.
- Contributed to the platform's embedded AI assistant while shipping daily features and larger multi-agent initiatives.
- Drove adoption of AI-assisted development with Cursor, Claude Code, and GitHub Copilot across developer pods.
- Shipped multi-agent workflows for a Fortune 500 tax enterprise using Azure OpenAI and custom MCP servers under enterprise governance and compliance constraints.
- Built interactive agentic AI demos used directly in IBM and Neudesic client engagements and proofs of concept.
- Benchmarked state management, tool integrations, and multi-agent handoffs that informed delivery standards.
- Drove GitHub Copilot adoption and AI-assisted development standards across internal software teams.
Selected projects
Selected Work
Three public projects with code, limitations, and enough evidence to inspect the implementation.
Open source · Python
Agent Autonomy Score
A CLI that scores implementation intent and code diffs, then recommends how closely a human should supervise the agent.
Open source · LLM routing
LLM Bidding System
An auction router where coding models bid on work and a supervised workflow records which model succeeded at what cost and risk level.
Open source · Python
Carry Trade Research Model
A finance-origin research system with FX collection, model experiments, a Flask API, and a React dashboard that labels missing and derived data honestly.
Other projects
- Event-Driven Multi-Agent ArchitectureArchitecture study · Multi-agent
A design study for running multi-agent AI systems on an event-driven service bus: agents subscribe to and emit events instead of calling each other directly, so they can be added, upgraded, or hot-swapped live without redeploying the rest of the system. Covers the event envelope, agent registry, reliability, and LLM cost on transports like Azure Service Bus and Kafka.
- Macro Oil Desk TerminalFull-stack · Azure AI Foundry
A trading-desk terminal for an oil desk — Brent/WTI dislocation z-scores, US inventory drawdown velocity, and AIS-based tanker fleet exposure by regulatory regime — with an Azure OpenAI and Microsoft Foundry-backed market-commentary panel and a real-time 3D Earth visualization. Built on Next.js and FastAPI, deployed on Azure.
- Autonomous Native AI Mobile Navigation AgentClaude Code · iOS
An iOS app for the Newport Beach Boat Show with an embedded multimodal assistant that interprets user intent and autonomously navigates the app's interface.
Writing
Technical notes from the work
Short, first-party notes on agent systems, reliability, and model economics.
The agentic future still runs on distributed systems
AI lowers the barrier to persistent, adaptive automation—and raises the reliability demands placed on the systems underneath it.
Read article →Multi-agent autonomy is compositional
A multi-agent system creates risks through the sequence of actions its participants can take together.
Read article →Model capability is converging. Model economics are not.
One real coding task cost roughly 80 times more with one frontier model than another, with negligible difference in the final result.
Read article →Capabilities
What I work with
Contact
Let's build something.
I'm open to conversations about agentic AI, enterprise automation, and engineering teams. LinkedIn is the best way to reach me.