Pilots that never become practice
A few engineers get great results. Everyone else gets inconsistent ones, and nobody can explain the difference or repeat it.
Consulting for engineering leaders
AI coding agents, adopted safely, affordably and at scale.
I help CTOs, VPs of Engineering and platform teams take Claude Code, Codex, OpenCode and Cursor from scattered experiments to a practice that is measurable, governed and owned by your engineers.
- 17+ years in software engineering and R&D leadership
- Led international teams of 25+ engineers
- Published on agent cost, workflow and security
# What ends up in the repo
instructions AGENTS.md, lean, tested
decisions docs/adr/ read by every agent
done one command, exits 0
mcp every server pinned
permissions scoped, sandbox by default
cost OTel → one ledger
guards stop on repeat failure
✓ reviewed like code
Problems I solve
Agents are easy to try and hard to scale.
Most organisations already have engineers using coding agents. The hard part is what happens next.
A bill nobody can explain
Usage is spread across tools, seats and API keys. Spend can't be attributed to a team, a repo or a model, so it can't be managed.
Instructions that sprawl and rot
AGENTS.md and CLAUDE.md files grow into long prose that is re-sent on every turn and still doesn't change how the agent behaves.
Agents that can touch too much
Unpinned MCP servers, broad permissions and no sandbox turn a helpful assistant into a supply-chain and data-access risk.
Review that can't keep up
More pull requests, same reviewers. “Tests pass” is not a finish line when the agent wrote the tests.
Sessions that loop, wait and stall
Agents poll long jobs, retry the same failing command or resume cold sessions, burning tokens and engineers' patience.
Services
Four ways I help, from first rollout to production.
Each one can stand alone or be combined into a single engagement.
Agentic coding adoption & enablement
Give every team the same, working foundation instead of a dozen private prompt collections.
- Instruction architecture: lean AGENTS.md / CLAUDE.md, linked workflows and skills, probe-tested in fresh sessions
- Decisions and specs in the repo (ADRs) so planner, builder and reviewer agents read the same page
- Team workflows: planning, a “definition of done” that runs as a command, review, session hand-offs
- Hands-on training in your own repositories, not toy examples
You leave with: A shared, versioned setup your teams can adopt and maintain themselves.
Cost & usage control
Know where the tokens go before deciding what to cut.
- Usage audits: context overhead, cache behaviour, cold restarts, idle polling
- One cost ledger from OpenTelemetry across Claude Code, Codex and other CLIs
- Model mix: which work needs a frontier model and which doesn't
- Subagent and session strategy that keeps context small and reviewable
You leave with: A baseline, a per-team view of spend, and a short list of changes ranked by impact.
Safe automation & governance
Let agents do more by deciding up front what they can touch.
- MCP server pinning and inventory, with a CI check that fails on unpinned servers
- Sandboxing and permission scoping per repository and environment
- Repeat-failure guards and hooks that stop loops instead of paying for them
- Practice for authorised, scoped security work with agents, such as patch verification and asset inventory
You leave with: Guardrails that live in tooling and CI, not in a policy document nobody reads.
Embedded engagement, prototype to production
Work alongside your platform or product team when the stakes are real.
- Forward-deployed with your engineers on a live initiative
- From first prototype to production rollout, including operating practices
- Hands-on architecture plus the people side: ownership, review and rollout
- Gradual hand-over so your team runs it without me
You leave with: A production-grade result and a team that owns it.
How engagements work
A clear path, in steps you can stop after.
Every engagement follows the same adoption path. You choose how far we go together and in which format.
Assessment
Find out where you really stand and what to do first.
- How it works
- Interviews, a look at your repos, configs and usage data where available.
- Outcome
- A written readout: usage and cost baseline, risk findings, and a prioritised roadmap.
Workshop
Build the foundation together with the engineers who'll use it.
- How it works
- Working sessions in your own repositories with one or more teams.
- Outcome
- Instruction files, ADR templates and guardrails your team built and understands.
Embedded advisory
Senior support while you roll out at scale.
- How it works
- A recurring cadence with your platform, DevEx or engineering leadership, plus hands-on work where it matters.
- Outcome
- Steady progress on the roadmap, reviews of what's working, course corrections as tools change.
Scope and terms are agreed in writing after an intro call. Get in touch to discuss what fits.
Why me · Approach
An engineering leader who still ships, not a slide deck.
Measure before you cut
Start from session logs, usage exports and telemetry. Opinions about “token waste” are cheap; a baseline isn't.
Put it in the repo, not in a chat
Instructions, decisions and definitions of done belong next to the code, versioned and reviewed like code.
Enforce with tooling, not prose
If a rule must always hold, it goes into a hook, a CI check or a pinned config, not another paragraph in AGENTS.md.
Tool-agnostic by design
Claude Code, Codex, OpenCode and Cursor all change monthly. The practices should survive a switch.
Hands-on, alongside your engineers
Recommendations are tested in your environment with your teams before they're rolled out.
Writing
Tested, practical write-ups on agentic coding.
How I think about cost, workflow and safety, in public. Most come with the commands and configs to try it yourself.
Security ·
PHP2Shell: Patch Today, Then Find the PHP You Forgot
How to check you're really patched, why most public “exploits” are junk, and how to get a coding agent to find every PHP version you run.
Read on Medium Usage audit ·Where Did My Claude Usage Go? Auditing Your Agent's Hidden Token Tax
Most “token leaks” aren't bugs. They're overhead paid on every turn. Measure it first, then cut it, one leak at a time.
Read on Medium Sessions ·Compact, Clear or Hand Off? Running Long Agent Sessions Without Burning Your Limit
Treat agent sessions as disposable and keep the state in files: when to compact, when to clear and when to write a hand-off.
Read on Medium Workflow ·Put Your Decisions in the Repo: ADRs as the Backbone of an AI Coding Workflow
Move specs and decision logs into the repository and wire them into AGENTS.md so planner, builder and reviewer agents read the same page.
Read on Medium Prompting ·Stop Asking Your AI for More. Ask It to Go Deeper.
Why “be more detailed” gets you longer answers rather than better ones, and the questions that change that.
Read on Medium Cost ·Cut Your Claude Code Bill by Up to 95%
Running Claude Code against a much cheaper Anthropic-compatible model, and what that does to the economics of long agentic sessions.
Read on Medium Cost ·Stop Burning Tokens: How Agent Orchestration and Knowledge Graphs Cut AI Coding Costs by 70%
Where agent credits actually go (re-establishing context, re-sending history) and how orchestration reduces it.
Read on MediumAB
About
Hands-on architecture, combined with people leadership.
I bring more than 17 years of experience leading software engineering, building complex software products and scaling engineering teams in regulated and data-intensive industries.
My background spans enterprise software, biotechnology, healthcare, automotive and scientific research. I've led software and data engineering for a biotech digital-twin platform, built safety and compliance tooling for automated-driving software, managed R&D software teams in healthcare and medical imaging, and led research software for digital rock analysis. Along the way I've contributed to more than 40 projects, spanning cloud platforms, data pipelines, AI, DevSecOps and scientific computing.
I've championed AI-assisted delivery and specification-driven development in engineering organisations. I've also taught computer science at university level and mentor students in online data and AI engineering programmes. I hold an M.Sc. in Computer Science and an MBA in Big Data & Analytics, and work in English, German, Spanish and Portuguese.
- Claude Code
- Codex
- OpenCode
- Cursor
- MCP
- OpenTelemetry
- Azure
- AWS
- CI/CD & DevSecOps
- Python
- .NET / C#
- C++
Which tools do you work with?
Claude Code, Codex, OpenCode and Cursor, plus the MCP servers, CI pipelines and telemetry around them. The recommendations are deliberately tool-agnostic: repo-resident instructions, pinned dependencies and measured usage carry over when you change or mix tools.
Is the goal to need fewer engineers?
No. The goal is that engineers who use agents get reliably good results, at a cost you can explain, without new security exposure. Where time is saved, you decide what to do with it.
How do you handle our code and data?
We agree it before any work starts. The default is to work inside your environment, under your NDA and access rules, with access scoped to the engagement. Security testing is only ever done with written authorisation and an agreed scope.
What do we need to get started?
A short intro call. For an assessment: a sponsor, one or two teams willing to be interviewed, and read access to configs and usage data where it exists. No preparation work is needed beyond that.
Remote or on-site?
Remote by default, with on-site workshops by arrangement. Sessions can run in English, German, Spanish or Portuguese.