I build AI systems that
earn autonomy.
Production LLM systems for agentic workflows, retrieval, evals, and progressive automation — built around real workflows, explicit trust, and measurable delegation.
- 50+ AI products shipped
- 150M+ User scale
- ~$20M Revenue impact
I do not resist automation. I sequence it.
Open to Principal / Staff AI Engineering roles and selective consulting · remote · EU/CET
Automation matures from real work.
I start by mapping how work actually happens — decisions, handoffs, exceptions, shortcuts, failure modes. I make state and trust explicit. Then I automate the parts that have earned enough evidence to be safely delegated.
The goal is not permanent control. It is a system where control can be safely transferred.
Map the real flow
The truth is in how work already happens, not in the spec.
Make trust explicit
State, evidence, evals, traces, permissions, and fallback behavior.
Delegate as evidence accumulates
Automation expands only where quality is measurable.
What I build
Agentic workflow systems
State machines, tool contracts, approvals, recovery paths.
Trust-critical retrieval
Grounding, permissions, citations, calibrated uncertainty.
Evaluation & observability
Traces, regression gates, delegation criteria.
Progressive automation
Manual → assisted → bounded delegation → trusted automation.
Across these projects, the pattern is the same: make the flow explicit, define trust boundaries, add evaluation, and automate progressively.
Content Ingestion for AI-Native Learning — SQL Walkthrough
Bounded prototype walkthrough showing how partial SQL teaching material can be transformed into a structured, inspectable learning graph with critique, repair, clarification, and deterministic validation.
What this proves: AI-native systems need structured intermediate representations and deterministic validation before generation can be trusted.
↗LLM agents and production systems at GriffinAI
End-to-end architecture and delivery of production agentic LLM products — Transaction Execution Agent, Cardano Proposal Examiner, and multi-agent ops.
What this proves: high-risk AI workflows can be made explicit, measurable, and recoverable in production.
↗Knowledge Assistant — Trust-Critical Retrieval Across Fragmented Internal Knowledge
Permission-aware internal knowledge assistant focused on trust-critical retrieval, citations, access control, and calibrated uncertainty across fragmented company knowledge.
What this proves: retrieval becomes useful only when permissions, uncertainty, and freshness are part of the workflow — not afterthoughts.
How autonomy is earned
- 01
Observe
Capture how work actually happens.
- 02
Structure
Make state, decisions, handoffs, and risks explicit.
- 03
Assist
Reduce operator friction without hiding responsibility.
- 04
Delegate
Automate bounded steps where quality is measurable.
- 05
Raise
Humans move from execution to direction and system design.
What I believe
Five working theses behind the systems I build.
-
The real workflow is the source of truth.
Most AI systems are designed against an idealized version of the work, not how it actually happens. The cost shows up in the second iteration, not the demo.
-
Human-in-the-loop is a stage, not an architecture.
Teams that treat HITL as the destination keep humans clicking approve forever. The point is to move them up the stack.
-
Evals are not QA.
They are the mechanism by which a system earns its next level of autonomy. Treating evals as a quality gate underuses them.
-
Trust is not a feeling. It is a measurable property.
State, uncertainty, evidence, and failure modes either exist explicitly in the system, or they show up later as production risk.
-
Automation is the outcome, not the starting point.
Production AI matures: observe → assist → delegate → trusted automation. Skipping stages is one of the most common ways production AI systems fail.
Current work
-
Building — Exocortex
A personal OS for knowledge work: one intake, one command surface, structured memory, and progressive automation. The same thesis I apply professionally, applied to my own workflow.
exocore.cx → -
Thinking — Engine vs Bolt-on
Why the real frontier in AI for learning is the delivery loop — not a chatbot beside static content.
Read essay →
Work with me
If you're building AI inside real product workflows — agents, retrieval, evals, observability, or progressive automation — I'm open to Principal / Staff AI engineering roles and selective technical collaboration.
Remote-first · Europe / CET-friendly · hands-on architecture and delivery.