Building AI agents for real-world workflows
I've spent my career helping people operate in difficult environments, often ideating and implementing solutions from the ground up. Today I apply the same principles, that enable people to work in environments with high uncertainty, to building AI agents.
- Y Combinator W21
- Founder, Greywing (2019)
- Founder, Citadel Maritime (2011)
- Former Royal Marine
Singapore
Building maritime crew change agents in Singapore
Drop me a note
Send a short message straight to Nick. Choose Email, WhatsApp, or Telegram for the reply.
At the moment I am building an end to end agent to manage crew changes for commercial vessels operating worldwide.
A crew change sounds simple. Then you see the vessels, masters, port agents, manning agencies, visas, flights, documents and approvals. You might be running a workflow with ~35 stakeholders each of whom has a different requirement. State is constantly changing and balancing human-in-the-loop approval gates with agentic automation is a delicate balancing act.
My work turns that difficult operational reality into an AI-agent workflow: structured, observable, secure, and human-controlled where it matters.
An operator-builder who has spent a career solving hard problems
I'm a former Royal Marine, repeat Founder, and AI-agent builder based in Singapore.
I really like solving hard problems.
I started my career in difficult operating environments, including military service and later maritime security work across piracy-risk regions such as Somalia, the Indian Ocean, West Africa, and Nigeria. I founded Citadel Maritime in 2011 and built it around protecting commercial shipping, superyachts, offshore energy projects, and high-value assets.
That work taught me that operational success depends on more than bravery or effort. It depends on coordination, reliable information, trusted suppliers, clear communication, and good judgement under pressure.
In 2019 I joined Entrepreneur First in Singapore, which unlocked my entry into technology. I met Hrishi Olickel there, and we co-founded Greywing later that year. When COVID hit, Greywing pivoted into crew-change automation as seafarers became stranded around the world and shipping companies had to navigate constantly changing port, travel, and immigration restrictions.
Today I'm focused on the next step: AI agents that operate inside real workflows. Not generic chatbots. Not dashboards with AI bolted on top. Agents that read documents, communicate over email, call tools, coordinate suppliers, track operational state, evaluate options, and bring humans in for approval and judgement.
An AI agent that can run a crew change end-to-end
A real crew change is not a form submission. It involves vessels, masters, crew managers, manning agencies, port agents, travel providers, visa rules, flights, costs, emails, documents, approval chains, and constant changes in state. I'm building agents that participate in that workflow directly, and keep a human in the loop where accountability matters.
Secure deployment
It's not complex to build an AI agent. The complexity comes when you deploy it safely inside real-world operations.
Crew Change Agent
An agentic workflow that coordinates a crew change across vessels, masters, manning agencies, port agents, travel providers, flights, visas, and approvals, from request to completion.
Email-native operations
The agent works through the channel maritime already runs on: email. It sends requests, receives replies, correlates responses, and updates case state.
Supplier coordination
It requests information and pricing from port agents and travel providers, compares responses, and structures them into something a crew manager can act on.
Document intelligence
PDFs, scans, crew lists, Statements of Facts, attachments. Turning messy documents into reliable data is a core input layer for the agent.
MCP and tool use
The agent connects to external tools and data: flight pricing, databases, operational systems, structured APIs, and workflow services.
Human approval loops
The goal isn't uncontrolled autonomy. It's reliable automation with clear escalation, approvals, auditability, and human accountability.
Observability and state
A serious agent is stateful and observable: you can see what it did, what it's waiting for, what evidence it used, and what needs approval.
Moving agentic AI from prototype to operational reality
The core question I keep coming back to: can an AI agent reliably run a real workflow where the inputs are messy, the communication happens over email, the data lives in documents and databases, and the result has to be commercially useful?
The crew-change agent is the clearest expression of that. A user can start a case and watch an agent send emails, receive replies, make decisions, update state, and progress the workflow in real time, built on agent frameworks, durable orchestration, structured tools, email infrastructure, stateful databases, streaming updates, and human approval points.
A lot of it depends on document intelligence: pipelines that use OCR, LLM classification, vision models, and structured extraction to turn PDFs and attachments into data an agent can act on. And on evaluation: how foundation model providers perform against real maritime tasks, and how to compare, route, and govern outputs when accuracy matters.
The supporting work, infrastructure, security, migrations, webhook validation, tenant boundaries, service accounts, dashboards, isn't the headline. But it's what turns an impressive demo into something that can operate safely near a real business process.
Real-world workflows inside legacy businesses
Workflows that break under uncertainty
The problems that interest me are the ones that look simple from the outside, then become fragile once weather, ports, people, documents, suppliers, timing, approvals, and incomplete information enter the room.
Problem pattern: Decisions with imperfect information, high coordination cost, and real consequences when the workflow stalls.
Legacy businesses with operational weight
I am interested in companies where the work is still carried by email, documents, phone calls, trust, local knowledge, and people who know how the operation actually runs.
Problem pattern: The workflow exists, but it is trapped across inboxes, spreadsheets, PDFs, supplier networks, and human memory.
Maritime coordination problems
Crew changes, port agents, vessel operations, travel providers, manning agencies, masters, visas, flights, and immigration rules are exactly the kind of operating environment where software has to respect reality.
Problem pattern: Many parties, shifting constraints, legacy systems, and a constant need to know the current state.
Agents that do operational work
I am not interested in generic chatbots. I am interested in agents that can read the documents, send the emails, call the tools, track the state, explain what happened, and bring a human in when judgement is needed.
Problem pattern: Turning an AI demo into a controlled workflow participant with state, tools, evidence, and approval gates.
Document and email-heavy work
The problems I keep coming back to involve unstructured inputs: PDFs, scans, crew lists, RFQs, replies, attachments, and email threads that contain the actual operational truth.
Problem pattern: Converting messy communication into structured data an agent can safely act on.
AI reliability in real tasks
I am interested in where models fail, where they are useful, and how to compare them against real business tasks instead of assuming one frontier model is always the answer.
Problem pattern: Evaluation, critique loops, routing, observability, and knowing when the system should stop and ask.
AI-native operations
Since 2024, I have become increasingly interested in how small teams can use AI agents not just inside the product, but inside maintenance, development, security review, migrations, and day-to-day operations.
Problem pattern: Re-architecting a business to do more with fewer people without losing control or accountability.
Problems worth comparing notes on
The people I want to speak to are usually looking at an operational workflow and asking: why is this still so manual, why is state so hard to see, and what would it take for an agent to help without creating new risk?
Problem pattern: Legacy workflows with enough complexity, repetition, and commercial value to justify serious agentic automation.
From the front line to the front of agentic AI
2003-2006
Recruitment and talent sourcing
Worked in recruitment, building a solid foundation in sales and go-to-market. Learned to read the skills gaps inside organisations and source the right person to solve a specific problem.
2007-2010
British Royal Marines
Served in the Royal Marines, including in Afghanistan & Sierra Leone. Learned to operate under pressure, coordinate in difficult environments, and make decisions where timing, trust, and clarity matter.
2011
Maritime security at sea
Left the Royal Marines and started working on board commercial vessels as part of armed security teams protecting ships from Somali piracy in the Indian Ocean and Gulf of Aden.
2011
Maritime security business development
Moved from vessel work into business development for a maritime security company. Started learning how commercial shipping buyers evaluated risk, trust, insurance, supplier reliability, and operational credibility.
2011
Building Citadel Maritime
Founded Citadel Maritime and spent the rest of the year putting the company architecture in place: trade control licence applications with the UK government, overseas weapons transfer and control processes, supplier networks, compliance, operating procedures, and early business development.
2012
First customers and offshore energy work
Started winning Citadel Maritime's first customers and moved from company setup into live operations. The work included support for Japan Oil and Gas Exploration Corporation activity in the Seychelles, bringing maritime security into offshore energy operations.
2013
Scaling Citadel Maritime
Grew the business through its second year as Citadel Maritime supported more complex customer work. The company also supported a Total Offshore Oman oil and gas exploration operation, extending from vessel protection into offshore energy security.
2014
From startup to operating business
By the third year, Citadel Maritime had become a real operating business. The work required hiring, supplier coordination, risk pricing, operational oversight, customer trust, and repeatable delivery across high-risk maritime environments.
2014
Superyachts and high-value assets
Expanded into superyacht security, attending the Monaco Yacht Show and protecting high-value vessels as they transited piracy-risk areas towards the Maldives, Seychelles, and the wider Indian Ocean.
2018
FPSO Firenze passage
Architected and managed the security plan for the FPSO Firenze passage across the Indian Ocean and through the Gulf of Aden to its refit in Dubai. The operation involved chartering two security escort vessels, deploying a security team, and providing security-management overwatch.
2019
Entrepreneur First and Greywing
Joined Entrepreneur First in Singapore through the EFSG5 Cohort, met Hrishi Olickel, and co-founded Greywing later that year. Greywing began from the idea of a maritime security intelligence platform that could turn operational risk data into better decisions.
2020
COVID crew-change crisis
When COVID disrupted global crew changes, Greywing re-architected the maritime security platform around where and when crew changes could safely take place. The product combined port-state control, immigration restrictions, local permissions, vessel tracking, and route-optimisation analysis.
2021
Y Combinator W21
Greywing joined Y Combinator's Winter 2021 batch, raised seed funding, and settled into a crew-change optimisation product for shipping teams navigating port, flight, travel, document, and immigration complexity.
2022-2023
Operational software for shipping
Continued moving Greywing from maritime risk intelligence into operational workflow software: helping crew managers structure messy communication, evaluate ports and routes, coordinate suppliers, and make decisions with better operational data.
2024
Breakeven and AI-native operations
After investor money ran out, re-architected Greywing to operate on a breakeven basis. More engineering work moved directly onto Nick, which accelerated the use of AI-assisted and agent-based workflows for maintenance, development, security review, and internal operations.
2025
Agent-based development and workflow automation
Built deeper agent-based workflows around maritime communication, port-agent outreach, email overwhelm, document intelligence, model evaluation, observability, and operational state. The work shifted from adding AI to software towards using agents to execute real parts of a workflow.
2026
Integrated maritime agents
The current focus is the next iteration of Greywing: moving from user-based vertical software workflows to fully integrated maritime agents that can run crew-change work end to end across email, documents, supplier coordination, flight pricing, state tracking, observability, and human approval.
Let's talk about applying AI agents to real operational work
I am fascinated with building AI agents.
Figuring out the intricacies of a complex problem and creating the systemic conditions for success, so an agent can execute that workflow repeatedly, is what I spend most of my time thinking about.
If you think about that too and are on a similar journey, maybe in a different industry, or if you're interested in what an AI agent would look like in your world, then we should connect.
Book a conversation