Where does work get stuck?
Look at repeated decisions, manual handoffs, delays, and the information people need to move forward.
AI transformation & production AI
AI transformation that gets implemented.
I help companies find the right workflows for AI, design practical solutions, and turn them into reliable production systems.
From discovery and prioritization through implementation and production reliability.
BeforeAI workflow
mailEmaildescriptionDocsgroupCRMgrid_onSheetsmanual handoffs
CaptureOne reliable intake
AI reasonsUsing your context
Tools executeAcross your systems
ReviewYour team stays in control
auto_awesome
Less busywork, fewer handoffs.Your team can inspect the workflow and improve it over time.
15+ yrsBuilding software
Still codingHands-on delivery
verified Strategy backed by delivery
Delta Labs // Multi-agent systemsGoogle // AI support agentNotamify // AI product & exitAWS // Platform reliability
Start with the work
Where would AI actually help?
Good transformation starts with a workflow and a measurable problem, before choosing a model or tool.
What is worth changing first?
Weigh expected value against data access, risk, integration effort, and the team’s readiness.
How will we know it works?
Set useful measures and human controls before delivery, then check them with real users.
Services
From opportunity to operation
Each stage produces something concrete to guide the next decision.
[01]Discover
Map the workflow, identify useful opportunities, and rank them by value and feasibility.
[02]Design
Define the experience, data, controls, and measures of success with the people who will use it.
[03]Deliver
Build and integrate the system, test it against real work, and put it in users’ hands.
[04]Scale
Measure quality, cost, and adoption; improve the workflow and support it in production.
Strategy + engineering
Strategy that can actually be implemented.
I help leaders select the right opportunity and work with teams to put it into production. The same delivery perspective informs the early decisions about data, integration, human review, and what success looks like.
Sometimes the right first step is a focused discovery engagement. Sometimes there is already a promising workflow ready to build or a live system that needs to work better.
How I work →Case study · Google
A support agent I led at Google
Deployed workflowTechnical support automation
Preparing responses for more than 70% of support inquiries.
Support specialists read each email thread, checked product systems and documentation, then wrote the reply. I led the project that automated most of that work. The support team still reviewed every response and decided what to send.
70%+Inquiries with review-ready replies
1 dayCustomer response time
ClearedGrowing inquiry backlog
verified_userSupport stayed in control. Specialists reviewed each response and decided what to send.
mail
Email threadCustomer inquiry
account_tree
ADK agentCheck issue · prepare reply
databaseProduct systemsmenu_bookDocumentation
fact_check
Validation layersPrivacy and quality checks
person_check
Review & sendTechnical support remains in control
Experience
Hi, I’m Michał.
Michał Pasierbski — AI Transformation & Production AI Consultant
I’ve spent more than 15 years building software and leading teams. At Google, I led the AI support-agent project above. Today I lead engineering for a multi-agent AI platform at Delta Labs, owning architecture and delivery while still working in the code.
That combination helps me connect a business opportunity to the technical and operational decisions needed to make it real.
Michał PasierbskiAI Transformation & Production AI Consultant
Technical depth
Built to work in production
Agent architecture, integration, evaluation, and observability matter when a useful idea becomes a dependable system.
account_treeFrameworks & Runtime
Google ADKLangGraphLlamaIndexPythonFastAPI
hubProtocols & Tooling
MCP ServersSandboxed ExecOpenAPIAG-UI
databaseVector & Memory
Redis VectorpgvectorPineconeQdrant
verified_userEvals & Guardrails
DeepEvalRagasLangfuseBraintrust
smart_toyFoundation Models
ClaudeOpenAIGeminiDeepSeek
Writing
Ideas for leaders. Details for builders.
I write about choosing useful AI work, making adoption practical, and the engineering decisions behind reliable production systems. The existing articles go deep on implementation.
Coding Agents01
Tripple H stack
Herdr runs the coding agents, Hunk gives me a place to review with them, and Heft keeps each task in its own workspace. A small stack for AI development.
Coding Agents02
Heft: Give Every Coding Agent Task Its Own Workspace
I built Heft to turn a task into a Git branch, worktree, and Herdr workspace with one command. Here is how it fits the way I work with coding agents.
Agent systems03
ADK sub_agents vs AgentTool: Two Delegation Patterns, One Critical Difference
Both sub_agents and AgentTool let a parent delegate to a child in Google ADK, but they have fundamentally different control flow semantics. Pick the wrong one and you get an agent that never hands back control - or a parent synthesizing output the child should have owned.
Agent systems04
Bridging Agents to UI: How AG-UI and CopilotKit Make It Stupidly Easy
Connecting your agent backend to a real-time UI used to mean weeks of custom streaming code. AG-UI and CopilotKit collapse that into a protocol and a few React components.
Ready to find a useful place for AI?
Bring the workflow, the opportunity, or the production problem you are working through.