I build AI agents that take routine work off teams.
I love turning a team’s repetitive work into systems that run on their own, with people stepping in only where they’re needed. My agents answer customers, take action in business systems and hand over to a person with full context.
From business school to building the systems behind the business.
I build conversational AI that does real work: agents that answer customers, look up their data, take action in business systems and hand over to a person with the full context when a case needs one. That’s been my daily work at Tiledesk for four years, across web chat, WhatsApp, voice and SMS.
I own each project end to end. I sit with the team to find the most costly routine work, agree on scope and KPIs, build, go live, then review real conversations every week to improve resolution rate and fix what the agent got wrong. On the Microsoft side I build the same thing with Copilot Studio, Power Automate, Power Apps and Dataverse.
A management engineering background means I start from the process, not the tool. I explain trade-offs clearly to non-technical teams and measure success by whether people keep using what I built.
Four areas I work in
Conversational AI and AI agents
- Multi-agent and sub-agent design
- Multi-channel: web chat, WhatsApp, voice, SMS
- Human handover with full context
- Agent instructions, procedures and guardrails
- Agents that take action through APIs and MCP
- Multilingual agents
Support AI operations
- Resolution rate, handovers and unanswered-question tracking
- Weekly conversation review and root-cause fixes
- Knowledge base and RAG configuration
- Helpdesk handover with AI ticket summaries
- Support data turned into feedback for Product
- Testing LLMs on quality, speed and cost before go-live
Microsoft Power Platform
- Copilot Studio agents and Agent Flows
- Power Apps canvas and model-driven
- Power Automate with error handling
- Dataverse data modelling
- Power Fx and AI Builder
Automation, integration and delivery
- Workflow automation (n8n, Make, Power Automate)
- API, webhook and OAuth integrations
- CRM lifecycle automation (HubSpot, Dynamics 365, Salesforce)
- Discovery and requirements translation
- Documentation and training
- Adoption support
Tools and platforms
Selected work, documented end to end
Each one walks through the problem, the design decisions and the result.
Multi-agent Support for a Payments Fintech
Agents and sub-agents across merchant, consumer app and general support on web and WhatsApp. They look up the customer’s own case through backend APIs and escalate when needed. 90%+ of conversations automated.
Read the case study→High-volume Retail Support with Helpdesk Handover
Around 100K conversations a month for a large home improvement retailer. Almost 80% of first-level requests automated, the rest handed to Zendesk as tickets with an AI-written title and summary.
Read the case study→Email AI Agent on Zendesk
A prototype on a fictional brand, built to learn Zendesk AI agents for email: two generative procedures, a mock account API in n8n and escalation routing. Six lessons documented.
Read the case study→Employee Leave Request Agent
A conversational agent that turns leave requests into a structured process. One-day requests auto-approve, multi-day requests go to the manager through Approvals, and every request lands in Dataverse.
Read the case study→
Field Technician Reporting App
Seven Dataverse tables, a four-screen canvas app with checklists that adapt to equipment type, and a flow that writes an AI summary and emails the right manager. No manual steps after submit.
Read the case study→B2B Product Catalog App
A tablet app for field sales: searchable catalog, per-product order history and an order form that creates the order and reduces stock together. Authored with Claude Code and the canvas-apps MCP plugin.
Read the case study→
VIP Loyalty Discount System
A server-side Power Fx function checks payment history, spend and order value, then a flow emails the discount code when a customer qualifies. One rule, reused anywhere it’s needed.
Read the case study→Measured outcomes from client work
Other client work
Internal support assistant
Tiledesk’s own support assistant answers from platform knowledge, hands over with context, and logs every unresolved issue with a priority and an assigned operator.
Website assistant and demo booking
Answers on web and WhatsApp, asks for the use case first so poor-fit visitors aren’t booked, and hands qualified leads to a second agent that books the slot. Hot leads go to sales with a summary for Customer Success.
WhatsApp support agent
Handles order tracking, refunds, address changes and product swaps through the client’s order management API.
Multilingual technician agent
For an industrial machinery manufacturer. Technicians describe issues or send photos on WhatsApp, the agent searches manuals and hands off to support with full context when needed.
Copilot Studio multi-channel agent
Copilot Studio agent with Tiledesk as the orchestration and handover layer. Context is kept on every escalation to a live agent.
AI sales advisor
Personalized recommendations in conversation from the live Shopify catalog through the Shopify MCP server.
Internal leave agent on two stacks
Copilot Studio and Dataverse for a Microsoft client; Tiledesk and n8n for Tiledesk’s own team, with no extra licences.
Mortgage scenario agent
An agent that simulates mortgage scenarios for a finance team.
Power Apps e-commerce back office
Canvas app on Dataverse for product and order data, flows for order processing, and an assistant that loads order history at the start of each chat.
Real-time free delivery check
A Dataverse Power Fx function checks the order subtotal against the threshold. The canvas app just calls it, so the rule stays reusable.
Dynamics 365 lead capture via Dataverse MCP
Agent resolves inquiries and creates or updates leads in Dynamics 365 through the Dataverse MCP server, with Entra ID auth and duplicate checks.
Claude + Google Calendar scheduling
Claude checks Calendar availability over MCP in real time and negotiates a slot with the customer automatically.
Tiledesk and Zendesk integration
OAuth setup, webhook and ticket creation with urgency classification. Published as Tiledesk’s official integration guide.
HubSpot lead qualification
Scoring on role, company, industry, budget and location, plus HubSpot status changes that trigger WhatsApp outreach handled by AI.
Academic background
MSc Management Engineering
Market and Competitive Analyst
BSc Business Management
Awards
Academic projects
L’Oréal: eco-friendly packaging
Sustainability campaign plan with market research, competitor analysis and campaign design.
Hotel Hamilton: pandemic pivot
Turning vacant rooms into remote workspaces and the restaurant into a hospitality training facility.
Haier: digital business innovation
Smart home opportunity mapping for 2025 and customer journeys built on IoT and AI.
What I’m learning
Most of this comes straight from questions I got in interviews and gaps I noticed in client work.
Enterprise security for Power Platform
Entra ID, managed identities, DLP at scale, PII handling in Copilot Studio agents.
Vector databases and RAG internals
Chunking, embeddings, retrieval scoring, hands-on with Chroma and Qdrant.
AB-410 Intelligent Applications Builder
Microsoft certification combining AI and Power Platform.
AI agent evaluation framework
Deflection, accuracy on test sets, escalation rate, latency and cost per conversation.
Intercom Fin
Guidance, procedures and workflows.
CSAT and A/B testing for support AI
Measuring satisfaction on AI conversations and testing agent changes against each other.
LangGraph fundamentals
Multi-agent patterns at code level, to contribute to architecture discussions directly.
ALM for Power Platform
Azure DevOps pipelines, managed solutions, environment variables and connection references.
Zendesk AI agents for email
Built a working email agent with generative procedures, escalation triggers and a mock account API.
Copilot Studio grounding and orchestration
Generative vs classic orchestration, knowledge grounding, debugging answer accuracy.
Custom connectors and OAuth 2.0
Auth code vs client credentials, Postman-first testing, connectors as agent actions.
Multi-agent orchestration patterns
Parent-child agents, shared components and reusable action libraries.