Work with me
Architecture consulting on agent memory, knowledge graphs, and agent identity and
trust — from a principal engineer who has shipped them in production. Five ways
in, from a time-boxed review to an ongoing design partnership. Berlin, and
remote across Europe.
Start a conversation by email
01
Agentic memory architecture
Design partner, ongoing
Designing the memory an agent actually needs: episodic, semantic and
procedural layers, the knowledge-graph substrate underneath them, and the
extraction-consolidation pipeline that keeps the whole thing from rotting. Plus
retrieval that does more than nearest-neighbour — graph traversal, PageRank and
community detection fused into one query.
- A memory architecture covering storage, retrieval and consolidation.
- A prioritised implementation plan with trade-offs and evaluation criteria.
- Your agent forgets things it was told last week.
- Retrieval returns plausible chunks and the agent still cannot answer multi-hop questions.
- You have a vector store and have started to suspect that is not a memory.
02
Fractional architect
Ongoing, part-time
Principal-level architecture without a full-time principal hire. A fixed
rhythm of design reviews, architecture decisions and pairing with your tech leads,
for teams building agents, knowledge graphs or memory-heavy products who need
someone who has made these calls before — and written down why.
- Senior architecture ownership for your agent and data platform, on a regular cadence.
- Decision records, design reviews and mentoring that stay with the team after I leave.
- You are building agents and nobody on the team has designed agent memory before.
- Your tech leads are strong, but architecture decisions keep getting made by default.
- You need principal-level judgement now, ahead of a principal-level hire.
03
Agent identity & trust
Project or advisory
Identity for agents, and trust between them. DID methods — did:web,
did:webvh, did:webs — KERI, verifiable credentials, GLEIF vLEI, agent reputation
and trust registry design. Built from founding-engineer work on a production SSI
platform, including a COVID verifiable-credentials system taken from zero to a
live airport integration in four weeks.
- An identity and trust model for your agents and their integrations.
- Protocol recommendations, trust boundaries and an implementation plan.
- Your agents need to prove who they are to something that is not your own backend.
- You are designing a trust registry and want to know what breaks at scale.
- Someone said "just use OAuth" and you suspect that answer is incomplete.
04
Architecture review
Time-boxed, written findings
A senior read of an existing system, ending in a written document you can
hand to your board or your team — what is sound, what will hurt, what to do
first. Frequently the cheapest way to find out whether you need anything else on
this page.
- A written assessment of strengths, risks and architectural trade-offs.
- Prioritised recommendations and a walkthrough with your team.
- A system works but nobody can say why, or what happens next.
- You are about to commit to a data architecture you cannot easily undo.
- You need an outside opinion your team will actually respect.
05
Workshops & training
Multi-day, on-site or remote
Teaching engineering teams the material behind the books: agent
architecture, knowledge graphs as agent memory, and local-first and edge AI.
Built around your team's real systems and decisions rather than assembled for a
conference slot.
- Team sessions on agent architecture, knowledge graphs and local-first AI.
- Exercises and reference material tied to your team’s engineering questions.
- Your team is shipping agents by pattern-matching on blog posts.
- You want a shared vocabulary before a big architectural decision, not after.
- You would rather your engineers learned this from someone who has shipped it.
How it works
Include a short description of your system, the decision you face, and your
timeline in your email. We agree the deliverables, duration and cadence in
the written proposal before the work starts.
- A short call to work out what you actually need — often not what the enquiry said.
- A written proposal with scope, shape and price.
- The work, with whatever cadence of contact suits you.
No price list. The right number depends on scope, duration and what the work is
worth to you, and I would rather find that out on a call than guess in public.
Where this comes from
Founding engineer at Affinidi — self-sovereign identity, DID
and Sidetree protocol work, and a COVID verifiable-credentials platform taken
from zero to a live airport integration in four weeks.
Founding engineer at Mykin — a privacy-first personal AI,
with on-device semantic memory: a personal knowledge graph with vector search
and clustering, running on a phone.
Before that: Zalando, TomTom, LiveIntent. Teams of five to fifteen, eighteen
teams co-ordinated on one rollout, systems serving forty million customers,
and three million events a minute.
Today: open-source memory systems — TypedGraph, Tiramemsu, OxigraphDB — and
independent research on bitemporal graphs, so the advice comes from code that
runs.
The full CV → Tools I have shipped →
- 18 teams
- about 100 engineers co-ordinated for the Zalando loyalty-programme rollout
- 40M+
- customers served by Transaction Core, the order backend I architected and led
- 4 weeks
- from zero to a live airport integration for a COVID verifiable-credentials platform
- $2M
- investment round I contributed to by representing the architecture to investors
- 3M/min
- events through the data-feeder services I owned at LiveIntent